Jerusalem Studies in Philosophy and History of Science

Meir Hemmo
Orly Shenker Editors

Quantum,
Probability,
Logic
The Work and Influence
of Itamar Pitowsky
Jerusalem Studies in Philosophy and History
of Science

Series Editor
Orly Shenker, The Hebrew University of Jerusalem, The Sidney M. Edelstein
Center for the History and Philosophy of Science, Technology and Medicine,
Jerusalem, Israel
Jerusalem Studies in Philosophy and History of Science sets out to present state of
the art research in a variety of thematic issues related to the fields of Philosophy of
Science, History of Science, and Philosophy of Language and Linguistics in their
relation to science, stemming from research activities in Israel and the near region
and especially the fruits of collaborations between Israeli, regional and visiting
scholars.

More information about this series at http://www.springer.com/series/16087

Meir Hemmo • Orly Shenker
Editors

Quantum, Probability, Logic

The Work and Influence of Itamar Pitowsky
Editors
Meir Hemmo Orly Shenker
University of Haifa The Hebrew University of Jerusalem
Haifa, Israel Jerusalem, Israel

ISSN 2524-4248 ISSN 2524-4256 (electronic)

Jerusalem Studies in Philosophy and History of Science
ISBN 978-3-030-34315-6 ISBN 978-3-030-34316-3 (eBook)
https://doi.org/10.1007/978-3-030-34316-3

© Springer Nature Switzerland AG 2020

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of
the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology
now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors, and the editors are safe to assume that the advice and information in this book
are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or
the editors give a warranty, expressed or implied, with respect to the material contained herein or for any
errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional
claims in published maps and institutional affiliations.

This Springer imprint is published by the registered company Springer Nature Switzerland AG.
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Dedicated to the memory of Itamar Pitowsky

Photograph by: Lev Vaidman

Preface

Itamar Pitowsky (1950–2010) was an outstanding philosopher of science, especially

physics, and a prominent and loved member of the small community in this field.
This volume marks a decade since Pitowsky’s untimely death. The essays in it
address his work and influence on the contemporary scene in the philosophy of
physics. They have been written by leading philosophers of physics and mathemati-
cal physicists especially for this volume, engaging in various ways with Pitowsky’s
work. The topics of the essays range from the interpretation of quantum mechanics
and the measurement problem, quantum nonlocality and contextuality, the nature of
the quantum probability, probability logic, quantum computation, the computational
complexity of physical systems, and the mathematical aspect of physics. Thus, the
essays provide a broad perspective on the state of the art in philosophy of physics
as well as its evolution during the past decade, touching upon open questions and
issues that are under debates in contemporary thinking.1 The result is an overview
of the depth and profoundness of Pitowsky’s work.
Pitowsky’s unique approach to the foundations of modern physics, computation,
and proof theory combined mathematical and physical skills with philosophical
sensitivity. While, in his publications, he cautiously refrained from committing
himself explicitly to broad metaphysical and epistemological views, those were
clearly present in his work and expressed in discussions in the many lectures and
workshops in which he was present (the last one held in his honor in Jerusalem in
2008; the lectures given there, including Pitowsky’s very last paper, appeared in the
volume Probability in Physics edited by Yemima Ben-Menahem and Meir Hemmo
Springer, 2012). If we can point to one central idea, which is the gist of many of
his contributions, it is a down-to-earth emphasis on empirical data as the central,
and possibly only, compelling basis on which one ought to form one’s scientific
and philosophical creed. This point of departure can be gleaned from his views
on the foundations of quantum mechanics, statistical mechanics, computation and
its relation to physics, logic and probability, and general philosophy of science.

1 The essays in the volume are arranged in alphabetical order of the authors’ surnames.

vii
viii Preface

Needless to say, he combined all of these with original mathematical insights, a

combination that is manifest in all of his publications, beginning with his early
work on the nonclassical nature of quantum probability presented in his influential
and widely cited book Quantum Probability – Quantum Logic (Springer 1989)
and developed jointly with his former teacher and collaborator Jeffrey Bub, in a
paper published in his last year (Bub and Pitowsky 2010; a lecture on which can
be seen online at https://vimeo.com/4845944). In the words of Jeffrey Bub and
William Demopoulos (2010), Itamar Pitowsky’s work “profoundly influenced our
understanding of what is characteristically nonclassical about quantum probabilities
and quantum logic.”2
A list of Pitowsky’s publications is provided on p. xiii of this volume; and his
intellectual genealogy as it appeared in his website is on p. ix.
We wish to thank the authors who wrote their articles especially for this volume,
Jeff Bub for helping us in preparing it, and the reviewers who have made significant
contributions to the articles they have read.

Haifa, Israel Meir Hemmo

Jerusalem, Israel Orly Shenker

2 Bub,
J., Demopoulos, W: Itamar Pitowsky 1950–2010. Studies in History and Philosophy of
Modern Physics 41, 85 (2010).
Itamar Pitowsky: Academic Genealogy
and the Arrow of Time

As you go backward in your family tree, the number of ancestors grows exponen-
tially with the number of generations, which means that it is practically impossible
to list all of them. The solution is easy enough: from the large number of ancestors,
you choose one with some distinction. Then you list the people in the branch leading
from this person to you while ignoring the rest. The rest include a multitude of
anonymous people, a few crooks, and perhaps some outright criminals. (If you
go back far enough, you are likely to find them). This is how your yichus is
manufactured.
Academic genealogy, on the other hand, consists of a single branch or at most a
few. You simply list your PhD supervisor, then his/her PhD supervisor, and so on. It
turns out that there is a web site devoted to academic genealogy of mathematicians
from which some of the information below is taken and can be continued further
down the generations to the seventeenth century. So here is part of it:

1. Jeffrey Bub, PhD from London University, 1966

ix
x Itamar Pitowsky: Academic Genealogy and the Arrow of Time

2. David Joseph Bohm, PhD from the University of California, Berkeley, 1943

(© Mark Edwards)
3. J. Robert Oppenheimer, PhD from the University of Göttingen, 1927

(Los Alamos National Laboratory)

4. Max Born, PhD from the University of Göttingen, 1907
Itamar Pitowsky: Academic Genealogy and the Arrow of Time xi

5. Carl David Tolmé Runge, PhD from the University of Berlin, 1880

6a. Supervisor #1: Ernst Eduard Kummer, PhD from Martin-Luther-Universität

Halle-Wittenberg, 1831

6b. Supervisor #2: Karl Theodor Wilhelm Weierstrass, PhD from the University of
Königsberg, 1854
xii Itamar Pitowsky: Academic Genealogy and the Arrow of Time

And so on . . .
Going forward in time changes one’s perspective completely. The great Weier-
strass has tens of thousands of academic descendants, so academic genealogy is not
a very big deal either. However, academic trees, unlike family trees, grow at different
rates backward and forward. As more PhD’s graduate, the entropy increases.
Itamar Pitowsky: Selected Publications

Itamar Pitowsky: Resolution of the Einstein-Podolsky-Rosen and Bell paradoxes.

Physical Review Letters, 48, 1299–1302 (1982)
Pitowsky, Itamar: Substitution and truth in quantum logic. Philosophy of Science
49(3), 380–401 (1982)
Pitowsky, Itamar: Where the Theory of Probability Fails. PSA: Proceedings of the
Biennial Meeting of the Philosophy of Science Association, pp. 616–623 (1982)
Pitowsky, Itamar: The Logic of Fundamental Processes: Nonmeasurable Sets and
Quantum Mechanics. Ph.D. Dissertation, The University of Western Ontario,
Canada (1983)
Pitowsky, Itamar: Deterministic model of spin and statistics. Physical Review D 27,
2316–2326 (1983)
Pitowsky, Itamar: Quantum mechanics and value definiteness. Philosophy of Sci-
ence 52(1), 154–156 (1985)
Pitowsky, Itamar: On the status of statistical inferences. Synthese 63(2), 233–247
(1985)
Pitowsky, Itamar: Unified field theory and the conventionality of geometry. Philos-
ophy of Science 51(4), 685–689 (1984)
Bub, Jeffrey, Pitowsky, Itamar: Critical notice on Popper’s postscript to the Logic
of Scientific Discovery. Canadian Journal of Philosophy 15(3), 539–552 (1986)
Itamar Pitowsky: The range of quantum probabilities. Journal of Mathematical
Physics 27, 1556–1565 (1986)
Itamar Pitowsky: Quantum logic and the “conspiracy” of nature. Iyyun, The
Jerusalem Philosophical Quarterly 36 (3/4), 204–234 (1987) (in Hebrew)
Pitowsky, Itamar: Quantum Probability – Quantum Logic. Lecture Notes in Physics
321, Springer-Verlag (1989)
Itamar Pitowsky: From George Boole to John Bell: The origin of Bell’s inequality.
In: Kafatos, M. (ed.) Bell’s Theorem, Quantum Theory and the Conceptions of
the Universe, pp. 37–49, Kluwer, Dordrecht (1989)
Itamar Pitowsky: The physical Church thesis and physical computational complex-
ity. lyyun, The Jerusalem Philosophical Quarterly 39, 81–99 (1990)

xiii
xiv Itamar Pitowsky: Selected Publications

Pitowsky, Itamar: Bohm’s quantum potentials and quantum gravity. Foundations of

Physics 21(3), 343–352 (1991)
Itamar Pitowsky: Correlation polytopes their geometry and complexity. Mathemat-
ical Programming, 50, 395–414 (1991)
Pitowsky, Itamar: Why does physics need mathematics? A comment. In: Edna
Ullmann-Margalit (ed.) The Scientific Enterprise, pp. 163–167, Kluwer Aca-
demic Publishers (1992)
Clifton, Robert, Pagonis, Constantine, Pitowsky, Itamar: Relativity, quantum
mechanics and EPR. PSA: Proceedings of the Biennial Meeting of the
Philosophy of Science Association, pp. 114–128 (1992)
Pitowsky, Itamar: George Boole’s ‘conditions of possible experience’ and the
quantum puzzle. British Journal for the Philosophy of Science 45(1), 95–125
(1994)
Pitowsky, Itamar: Laplace’s demon consults an oracle: The computational complex-
ity of prediction. Studies in History and Philosophy of Modern Physics 27(2),
161–180 (1996)
Pitowsky, Itamar, Shoresh, Noam: Locality, factorizability, and the Maxwell-
Boltzmann distribution. Foundations of Physics 26(9), 1231–1242 (1996)
Itamar Pitowsky: Infinite and finite Gleason’s theorems and the logic of indetermi-
nacy. Journal of Mathematical Physics 39, 218–228 (1998)
Pitowsky, Itamar: Local fluctuations and local observers in equilibrium statistical
mechanics. Studies in History and Philosophy of Modern Physics 32(4), 595–
607 (2001)
Pitowsky, Itamar, Svozil, Karl: Optimal tests of quantum nonlocality. Physical
Review A 64, 014102 (2001)
Pitowsky, Itamar: Range theorems for quantum probability and entanglement.
arXiv:quant-ph/0112068 (2001)
Ben-Menahem, Yemima, Pitowsky, Itamar (2001). Introduction. Studies in History
and Philosophy of Modern Physics 32(4), 503–510 (2001)
Pitowsky, Itamar: Quantum speed-up of computations. Philosophy of Science 69
(S3), S168–S177 (2002)
Pitowsky, Itamar: Generalizations of Kochen and Specker’s theorem and the
effectiveness of Gleason’s theorem. Studies in History and Philosophy of Modern
Physics 35(2), 177–194 (2003)
Pitowsky, Itamar: Betting on the outcomes of measurements: A Bayesian theory of
quantum probability. Studies in History and Philosophy of Modern Physics 34
(3), 395–414 (2003)
Hemmo, Meir, Pitowsky, Itamar: Probability and nonlocality in many minds
interpretations of quantum mechanics. British Journal for the Philosophy of
Science 54(2), 225–243 (2003)
Shagrir, Oron, Pitowsky, Itamar: Physical hypercomputation and the church–turing
thesis. Minds and Machines 13(1), 87–101 (2003)
Pitowsky, Itamar: Macroscopic objects in quantum mechanics: A combinatorial
approach. Physical Review A 70, 022103 (2004)
Itamar Pitowsky: Selected Publications xv

Pitowsky, Itamar: Random witnesses and the classical character of macroscopic

objects. philsci-archive.pitt.edu/id/eprint/1703 (2004)
Hrushovski, Ehud, Pitowsky, Itamar: Generalizations of Kochen and Specker’s
theorem and the effectiveness of Gleason’s theorem. Studies in History and
Philosophy of Modern Physics 35(2), 177–194 (2004)
Pitowsky, Itamar: On the definition of equilibrium. Studies in History and Philoso-
phy of Modern Physics 37(3), 431–438 (2006)
Dolev, Shahar, Pitowsky, Itamar, Tamir Boaz: Grover’s quantum search algorithm
and Diophantine approximation. Physical Review A 73, 022308 (2006)
Pitowsky, Itamar: Quantum mechanics as a theory of probability. In: Demopoulos,
William, Pitowsky, Itamar (eds.) Physical Theory and Its Interpretation: Essays
in Honor of Jeffrey Bub. Western Ontario series in philosophy of science (72),
pp. 213–240, Springer (2007)
Pitowsky, Itamar: From logic to physics: How the meaning of computation changed
over time. In: Cooper, B.S., Löwe, B., Sorbi, A. (eds.) Computation and Logic in
the Real World, pp. 621–631, Springer, Berlin Heidelberg (2007)
Pitowsky, Itamar: On Kuhn’s The Structure of Scientific Revolutions. Iyyun: The
Jerusalem Philosophical Quarterly 56, 119 (2007)
Hemmo, Meir, Pitowsky, Itamar: Quantum probability and many worlds. Studies in
History and Philosophy of Modern Physics 38(2), 333–350 (2007)
Pitowsky, Itamar: Geometry of quantum correlations. Physical Review A 77,
062109 (2008)
Pitowsky, Itamar: New Bell inequalities for the singlet state: Going beyond the
Grothendieck bound. Journal of Mathematical Physics 49, 012101 (2008)
Bub, Jeffrey, Pitowsky, Itamar (2010). Two dogmas about quantum mechanics. In:
Saunders, Simon, Barrett, Jonathan, Kent, Adrian, Wallace, David (eds.) Many
Worlds? Everett, Quantum Theory, & Reality, pp. 433–459, Oxford University
Press (2010).
Pitowsky, Itamar: Typicality and the role of the Lebesgue measure in statistical
mechanics. In: Ben-Menahem, Yemima, Hemmo Meir (eds.) Probability in
Physics, pp. 41–58, Springer-Verlag Berlin Heidelberg (2012)
Contents

1 Classical Logic, Classical Probability, and Quantum Mechanics . . . . . 1

Samson Abramsky
2 Why Scientific Realists Should Reject the Second Dogma
of Quantum Mechanics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
Valia Allori
3 Unscrambling Subjective and Epistemic Probabilities . . . . . . . . . . . . . . . . . 49
Guido Bacciagaluppi
4 Wigner’s Friend as a Rational Agent . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
Veronika Baumann and Časlav Brukner
5 Pitowsky’s Epistemic Interpretation of Quantum Mechanics
and the PBR Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
Yemima Ben-Menahem
6 On the Mathematical Constitution and Explanation of Physical
Facts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
Joseph Berkovitz
7 Everettian Probabilities, The Deutsch-Wallace Theorem and
the Principal Principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
Harvey R. Brown and Gal Ben Porath
8 ‘Two Dogmas’ Redux. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
Jeffrey Bub
9 Physical Computability Theses. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
B. Jack Copeland and Oron Shagrir
10 Agents in Healey’s Pragmatist Quantum Theory: A
Comparison with Pitowsky’s Approach to Quantum Mechanics . . . . . 233
Mauro Dorato

xvii
xviii Contents

11 Quantum Mechanics As a Theory of Observables and States

(And, Thereby, As a Theory of Probability) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257
John Earman and Laura Ruetsche
12 The Measurement Problem and Two Dogmas About Quantum
Mechanics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 285
Laura Felline
13 There Is More Than One Way to Skin a Cat: Quantum
Information Principles in a Finite World . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 305
Amit Hagar
14 Is Quantum Mechanics a New Theory of Probability? . . . . . . . . . . . . . . . . . 317
Richard Healey
15 Quantum Mechanics as a Theory of Probability . . . . . . . . . . . . . . . . . . . . . . . . 337
Meir Hemmo and Orly Shenker
16 On the Three Types of Bell’s Inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 353
Gábor Hofer-Szabó
17 On the Descriptive Power of Probability Logic . . . . . . . . . . . . . . . . . . . . . . . . . 375
Ehud Hrushovski
18 The Argument Against Quantum Computers . . . . . . . . . . . . . . . . . . . . . . . . . . . 399
Gil Kalai
19 Why a Relativistic Quantum Mechanical World Must Be
Indeterministic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 423
Avi Levy and Meir Hemmo
20 Subjectivists About Quantum Probabilities Should Be Realists
About Quantum States . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 449
Wayne C. Myrvold
21 The Relativistic Einstein-Podolsky-Rosen Argument . . . . . . . . . . . . . . . . . . 467
Michael Redhead
22 What Price Statistical Independence? How Einstein Missed
the Photon . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 479
Simon Saunders
23 How (Maximally) Contextual Is Quantum Mechanics? . . . . . . . . . . . . . . . . 505
Andrew W. Simmons
24 Roots and (Re)sources of Value (In)definiteness Versus
Contextuality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 521
Karl Svozil
25 Schrödinger’s Reaction to the EPR Paper . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 545
Jos Uffink
Contents xix

26 Derivations of the Born Rule . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 567

Lev Vaidman
27 Dynamical States and the Conventionality of (Non-) Classicality . . . . . 585
Alexander Wilce
Contributors

Samson Abramsky Department of Computer Science, University of Oxford,

Oxford, UK
Valia Allori Department of Philosophy, Northern Illinois University, Dekalb, IL,
USA
Guido Bacciagaluppi Descartes Centre for the History and Philosophy of the
Sciences and the Humanities, Utrecht University, Utrecht, Netherlands
Veronika Baumann Vienna Center for Quantum Science and Technology (VCQ),
Faculty of Physics, University of Vienna, Vienna, Austria
Faculty of Informatics, Università della Svizzera italiana, Lugano, Switzerland
Yemima Ben-Menahem Department of Philosophy, The Hebrew University of
Jerusalem, Jerusalem, Israel
Joseph Berkovitz Institute for the History and Philosophy of Science and Technol-
ogy, University of Toronto, Victoria College, Toronto, ON, Canada
Harvey R. Brown Faculty of Philosophy, University of Oxford, Oxford, UK
Časlav Brukner Vienna Center for Quantum Science and Technology (VCQ),
Faculty of Physics, University of Vienna, Vienna, Austria
Institute of Quantum Optics and Quantum Information (IQOQI), Austrian Academy
of Sciences, Vienna, Austria
Jeffrey Bub Department of Philosophy, Institute for Physical Science and Technol-
ogy, Joint Center for Quantum Information and Computer Science, University of
Maryland, College Park, MD, USA
B. Jack Copeland Department of Philosophy, University of Canterbury,
Christchurch, New Zealand
Mauro Dorato Department of Philosophy, University of Rome Tre, Rome, Italy
John Earman Department of History and Philosophy of Science, University of
Pittsburgh, Pittsburgh, PA, USA
xxi
xxii Contributors

Laura Felline Department of Philosophy, University of Roma Tre, Rome, Italy

Amit Hagar Department of History and Philosophy of Science and Medicine,
Indiana University, Bloomington, IN, USA
Richard Healey Philosophy Department, University of Arizona, Tucson, AZ, USA
Meir Hemmo Department of Philosophy, University of Haifa, Haifa, Israel
Gábor Hofer-Szabó Research Center for the Humanities, Budapest, Hungary
Ehud Hrushovski Mathematical Institute, Oxford, UK
Gil Kalai Einstein Institute of Mathematics, The Hebrew University of Jerusalem,
Jerusalem, Israel
Efi Arazy School of Computer Science, Interdisciplinary Center Herzliya, Herzliya,
Israel
Avi Levy Department of Philosophy, University of Haifa, Haifa, Israel
Wayne C. Myrvold Department of Philosophy, The University of Western Ontario,
London, ON, Canada
Gal Ben Porath Department of History and Philosophy of Science, University of
Pittsburgh, Pittsburgh, PA, USA
Michael Redhead Centre for Philosophy of Natural and Social Science, London
School of Economics and Political Science, London, UK
Laura Ruetsche Department of Philosophy, University of Michigan, Ann Arbor,
MI, USA
Simon Saunders Faculty of Philosophy, University of Oxford, Oxford, UK
Oron Shagrir Department of Philosophy, The Hebrew University of Jerusalem,
Jerusalem, Israel
Orly Shenker Edelstein Center for History and Philosophy of Science, The Hebrew
University of Jerusalem, Jerusalem, Israel
Andrew W. Simmons Department of Physics, Imperial College London, London,
UK
Karl Svozil Institute for Theoretical Physics, Vienna University of Technology,
Vienna, Austria
Jos Uffink Philosophy Department and Program in History of Science, Technology
and Medicine, University of Minnesota, Minneapolis, MN, USA
Lev Vaidman School of Physics and Astronomy, Tel-Aviv University, Tel Aviv,
Israel
Alexander Wilce Department of Mathematical Sciences, Susquehanna University,
Selinsgrove, PA, USA
Chapter 1
Classical Logic, Classical Probability,
and Quantum Mechanics

Samson Abramsky

Abstract We give an overview and conceptual discussion of some of our results

on contextuality and non-locality. We focus in particular on connections with the
work of Itamar Pitowsky on correlation polytopes, Bell inequalities, and Boole’s
“conditions of possible experience”.

Keywords Logic · Probability · Correlation polytopes · Bell inequalities ·

Contextual fraction · Bell’s theorem

1.1 Introduction

One of Itamar Pitowsky’s most celebrated contributions to the foundations of quan-

tum mechanics was his work on correlation polytopes Pitowsky (1989, 1991, 1994),
as a general perspective on Bell inequalities. He related these to Boole’s “conditions
of possible experience” Boole (1862), and emphasized the links between correlation
polytopes and classical logic.
My own work on the sheaf-theoretic approach to non-locality and contextuality
with Adam Brandenburger Abramsky and Brandenburger (2011), on logical Bell
inequalities with Lucien Hardy Abramsky and Hardy (2012), on robust constraint
satisfaction with Georg Gottlob and Phokion Kolaitis Abramsky et al. (2013),
and on the contextual fraction with Rui Soares Barbosa and Shane Mansfield
Abramsky et al. (2017), is very much in a kindred spirit with Pitowsky’s pioneering
contributions. I will survey some of this work, making a number of comparisons and
contrasts with Pitowsky’s ideas.

S. Abramsky ()
Department of Computer Science, University of Oxford, Oxford, UK
e-mail: samson.abramsky@cs.ox.ac.uk

© Springer Nature Switzerland AG 2020 1

M. Hemmo, O. Shenker (eds.), Quantum, Probability, Logic,
Jerusalem Studies in Philosophy and History of Science,
https://doi.org/10.1007/978-3-030-34316-3_1
2 S. Abramsky

1.2 Boole’s “Conditions of Possible Experience”

We quote Pitowsky’s pellucid summary (Pitowsky 1994, p. 100):

Boole’s problem is simple: we are given rational numbers which indicate the relative
frequencies of certain events. If no logical relations obtain among the events, then the
only constraints imposed on these numbers are that they each be non-negative and less
than one. If however, the events are logically interconnected, there are further equalities or
inequalities that obtain among the numbers. The problem thus is to determine the numerical
relations among frequencies, in terms of equalities and inequalities, which are induced
by a set of logical relations among the events. The equalities and inequalities are called
“conditions of possible experience”.

More formally, we are given some basic events E1 , . . . , En , and some boolean
functions ϕ 1 , . . . , ϕ m of these events. Such a function ϕ can be described by a
propositional formula in the variables E1 , . . . , En .
Suppose further that we are given probabilities p(Ei ), p(ϕ j ) of these events.

Question: What numerical relationships between the probabilities

can we infer from the logical relationships between the events?

Pitowksy’s approach is to define a correlation polytope c(E,  ϕ ) induced by

the given events, and to characterize the “conditions of possible experience” as
the facet-defining inequalities for this polytope. He emphasizes the computational
difficulty of obtaining these inequalities, and gives no direct characterization other
than a brute-force computational approach. Indeed, one of his main results is the NP-
completeness of an associated problem, of determining membership of the polytope.
We shall return to these ideas later.
The point we wish to make now is that it is possible to give a very direct
answer to Boole’s question, which explicitly relates logical consistency conditions
to arithmetical relationships. This insight is the key observation in Abramsky and
Hardy (2012).
With notation as above, suppose that the formulas ϕ j are not simultaneously
m−1
satisfiable. This means that i=1 ϕ i → ¬ϕ m , or by contraposition and De
Morgan’s law:


m−1
ϕm → ¬ϕ i .
i=1

Passing to probabilities, we infer:


m−1 
m−1 
m−1 
m−1
p(ϕ m ) ≤ p( ¬ϕ i ) ≤ p(¬ϕ i ) = (1−p(ϕ i )) = (m−1)− p(ϕ i ).
i=1 i=1 i=1 i=1
1 Classical Logic, Classical Probability, and Quantum Mechanics 3

The first inequality is the monotonicity of probability, i.e. E ⊆ E  implies p(E) ≤

p(E  ), while the second is the subadditivity of probability measures.1 Collecting
terms, we obtain:


m
p(ϕ i ) ≤ m − 1. (1.1)
i=1

Thus a consistency condition on the formulas ϕ i implies a non-trivial arithmetical

inequality on the probabilities p(ϕ i ).

1.2.1 An Example: The Bell Table

We can directly apply the above inequality to deduce a version of Bell’s theorem
Bell (1964). Consider the following table.

A B (0, 0) (1, 0) (0, 1) (1, 1)

a b 1/2 0 0 1/2
a b 3/8 1/8 1/8 3/8
a b 3/8 1/8 1/8 3/8
a b 1/8 3/8 3/8 1/8

Here we have two agents, Alice and Bob. Alice can choose from the measurement
settings a or a  , and Bob can choose from b or b . These choices correspond to
the rows of the table. The columns correspond to the joint outcomes for a given
choice of settings by Alice and Bob, the two possible outcomes for each individual
measurement being represented by 0 and 1. The numbers along each row specify a
probability distribution on these joint outcomes. Thus for example, the entry in row
2, column 2 of the table says that when Alice chooses setting a and Bob chooses
setting b , then with probability 1/8, Alice obtains a value of 1, and Bob obtains a
value of 0.
A standard version of Bell’s theorem uses the probability table given above. This
table can be realized in quantum mechanics, e.g. by a Bell state

|00 + |11
√ ,
2

subjected to spin measurements in the XY -plane of the Bloch sphere, at a relative

angle of π/3. See the supplemental material in Abramsky et al. (2017) for details.

1 Also known as Boole’s inequality (Weisstein 2019).

4 S. Abramsky

1.2.1.1 Logical Analysis of the Bell Table

We now pick out a subset of the elements of each row of the table, as indicated in
the following table.

A B (0, 0) (1, 0) (0, 1) (1, 1)

a b 1/2 0 0 1/2
a b 3/8 1/8 1/8 3/8
a b 3/8 1/8 1/8 3/8
a b 1/8 3/8 3/8 1/8

We can think of basic events Ea , Ea  , Eb , Eb , where e.g. Ea is the event that
the quantity measured by a has the value 0. Note that, by our assumption that each
measurement has two possible outcomes,2 ¬Ea is the event that a has the value 1.
Writing simply a for Ea etc., the highlighted positions in the table are represented
by the following propositions:

ϕ1 = a∧b ∨ ¬a ∧ ¬b = a ↔ b
ϕ2 = a ∧ b ∨ ¬a ∧ ¬b = a ↔ b
ϕ3 = a ∧ b ∨ ¬a  ∧ ¬b = a ↔ b
ϕ4 = ¬a  ∧ b ∨ a  ∧ ¬b = a ⊕ b .

The first three propositions pick out the correlated outcomes for the variables they
refer to; the fourth the anticorrelated outcomes. These propositions are easily seen
to be contradictory. Indeed, starting with ϕ 4 , we can replace a  with b using ϕ 3 ,
b with a using ϕ 1 , and a with b using ϕ 2 , to obtain b ⊕ b , which is obviously
unsatisfiable.
We see from the table that p(ϕ 1 ) = 1, p(ϕ i ) = 6/8 for i = 2, 3, 4. Hence the
violation of the Bell inequality (1.1) is 1/4.
We may note that the logical pattern shown by this jointly contradictory family
of propositions underlies the familiar CHSH correlation function.

1.2.2 The General Form

We can generalize the inequality (1.1). Given a family of propositions  = {ϕ i },

we say it is K-consistent if the size of the largest consistent subfamily of  is K.3 If
a family {ϕ i }m
i=1 is not simultaneously satisfiable, then it must be K-consistent for
some K < m.

2 And, implicitly, the assumption that each quantity has a definite value, whether we measure it or
not.
3 The size of a family {ϕ }
j j ∈J is the cardinality of J . Note that repetitions are allowed in the family.
1 Classical Logic, Classical Probability, and Quantum Mechanics 5

Theorem 1 Suppose that we have a K-consistent family {ϕ i }m

i=1 over the basic

events E1 , . . . , En . For any probability distribution on the set 2E of truth-value
assignments to the Ei , with induced probabilities p(ϕ i ) for the events ϕ i , we have:


m
p(ϕ i ) ≤ K. (1.2)
i=1

See Abramsky and Hardy (2012) for the (straightforward) proof. Note that the
basic inequality (1.1) is a simple consequence of this result.
We thus have a large class of inequalities arising directly from logical consistency
conditions. As shown in Abramsky and Hardy (2012), even the basic form (1.1) is
sufficient to derive the main no-go results in the literature, including the Hardy, GHZ
and Kochen-Specker “paradoxes”.
More remarkably, as we shall now go on to see, this set of inequalities
is complete. That is, every facet-defining inequality for the non-local, or more
generally non-contextual polytopes, is equivalent to one of the form (1.2). In this
sense, we have given a complete answer to Boole’s question.
The following quotation from Pitowsky suggests that he may have envisaged the
possibility of such a result (Pitowsky 1991, p. 413):
In fact, all facet inequalities for c(n) should follow from “Venn diagrams”, that is, the
possible relations among n events in a probability space.

1.3 From Correlation Polytopes to Non-contextual Polytopes

We continue with the notation E1 , . . . , En for basic events, and ϕ 1 , . . . , ϕ m for

boolean combinations of these events. We write 2 := {0, 1}.
Pitowsky defines the correlation polytope c(E,  ϕ ) ⊆ Rn+m as follows. Each

assignment s ∈ 2 of 0 or 1 to the basic events Ei extends to a truth assignment
E

for the formulae ϕ j by the usual truth-table method, and hence determines a 0/1-
vector vs in Rn+m . The polytope c(E, ϕ ) is defined to be the convex hull of the set

of vectors vs , for s ∈ 2 .
E

In particular, Pitowsky focusses almost exclusively on the case where the

formulae are pairwise conjunctions ϕ ij := Ei ∧ Ej . One effect of restricting to
this case is that m is bounded quadratically by n. We shall return to this point when
we discuss complexity issues.
The main computational problem that Pitowsky focusses on is the following:

Instance: a rational vector v ∈ Qn+m .

Question: is v in c(E,ϕ
 )?
6 S. Abramsky

For the case where ϕ is the set of all pairwise conjunctions ϕ ij of basic events,
Pitowsky shows that this problem is NP-complete.
We can see that the (quite standard) version of Bell’s theorem described in
Sect. 1.2.1 amounts to the statement that the vector

v = [pa , pa  , pb , pb , 1, 6/8, 6/8, 6/8]T

is not in the correlation polytope c(Ea , Ea  , Eb , Eb , ϕ 1 , ϕ 2 , ϕ 3 , ϕ 4 ). Here pa is the

probability of the event Ea . This can be calculated as a marginal from the first or
second rows of the Bell table, as pa = 1/2. The fact that we get the same answer,
whether we use the first or second row, is guaranteed by the no-signalling condition
(Jordan 1983; Ghirardi et al. 1980), which will hold for all such tables which can
be generated using quantum states and observables (Abramsky and Brandenburger
2011). Similarly for the other basic events.
We make a number of comments:
• Firstly, notice that the Bell table as given in Sect. 1.2.1 corresponds to the data
arising directly from a Bell experiment. This and similar tables should be seen as
setting the standard for what we are trying to describe.
• From this point of view, we notice firstly that to get a natural correspondence
between these tables and correlation polytopes, we need to consider propositions
beyond pairwise conjunctions. This makes a huge difference as regards the size
of problem instances. While there are only O(n2 ) pairwise conjunctions on n
n
variables, there are 22 distinct boolean functions.
• At the same time, there is redundancy in the correlation polytope representation,
since thanks to the no-signalling condition, the probabilities of the basic events
can be calculated as marginals from the probabilities of the joint outcomes.
• The most important point is that there is a structure inherent in the Bell table
and its generalizations, which is “flattened out” by the correlation polytope
representation. Capturing this structure explicitly can lead to deeper insights into
the non-classical phenomena arising in quantum mechanics.
Taking up the last point, we quote again from Pitowsky’s admirably clear
exposition (Pitowsky 1994, p. 111–112):
One of the major purposes of quantum mechanics is to organize and predict the relative
frequencies of events observable in various experiments — in particular the cases where
Boole’s conditions are violated. For that purpose a mathematical formalism has been
invented which is essentially a new kind of probability theory. It uses no concept of
‘population’ but rather a primitive concept of ‘event’ or more generally ‘observable’ (which
is the equivalent of the classical ‘random variable’). In addition, to every particular physical
system (which can be one ‘thing’ — an electron, for example — or consists of a few
‘things’) the theory assigns a state. The state determines the probabilities for the events or,
more generally, the expectations of the observables. What this means operationally is that
if we have a source of physical systems, all in the same state, then the relative frequency of
a given event will approach the value of the probability, which is theoretically determined
by the state.
1 Classical Logic, Classical Probability, and Quantum Mechanics 7

For certain families of events the theory stipulates that they are commeasurable. This
means that, in every state, the relative frequencies of all these events can be measured on
one single sample. For such families of events, the rules of classical probability — Boole’s
conditions in particular — are valid. Other families of events are not commeasurable, so
their frequencies must be measured in more than one sample. The events in such families
nevertheless exhibit logical relations (given, usually, in terms of algebraic relations among
observables). But for some states, the probabilities assigned to the events violate one or
more of Boole’s conditions associated with those logical relations.

The point we would like to emphasize is that tables such as the Bell table in
Sect. 1.2.1 can—and do—arise from experimental data, without presupposing any
particular physical theory. This data does clearly involve concepts from classical
probability. In particular, for each set C of “commeasurable events”, there is a
sample space, namely the set of possible joint outcomes of measuring all the
observables in C. Moreover, there is a well-defined probability distribution on this
sample space.
Taking the Bell table for illustration, the “commeasurable sets”, or contexts, are
the sets of measurements labelling the rows of the table. The sample space for a
given row, with measurement α by Alice and β by Bob, is the set of joint outcomes
(α = x, β = y) with x, y ∈ {0, 1}. The probabilities assigned to these joint
outcomes in the table form a (perfectly classical) probability distribution on this
sample space.
Thus the structure of the table as a whole is a family of probability distributions,
each defined on a different sample space. However, these sample spaces are not
completely unrelated. They overlap in sharing common observables, and they “agree
on overlaps” in the sense that they have consistent marginals. This is exactly the
force of the no-signalling condition.
The nature of the “non-classicality” of the data tables generated by quantum
mechanics (and also elsewhere, see e.g. Abramsky 2013), is that there is no
distribution over the global sample space of outcomes for all observables, which
recovers the empirically accessible data in the table by marginalization. This is
completely equivalent to a statement of the more traditional form: “there is no local
or non-contextual hidden variable model” (or, in currently fashionable terminology:
there is no “ontological model” of a certain form).
Our preferred slogan for this state of affairs is:

Contextuality (with non-locality as a special case) arises when we have a

family of data which is locally consistent, but globally inconsistent.

1.3.1 Formalization

We briefly summarise the framework introduced in Abramsky and Brandenburger

(2011), and extensively developed subsequently. The main objects of study are
empirical models: tables of data, specifying probability distributions over the joint
8 S. Abramsky

outcomes of specified sets of compatible measurements. These can be thought of as

statistical data obtained from some experiment or as the observations predicted by
some theory.
A measurement scenario is an abstract description of a particular experimental
setup. It consists of a triple X, M, O where: X is a finite set of measurements; O is
a finite set of outcome values for each measurement; and M is a set of subsets of X.
Each C ∈ M is called a measurement context, and represents a set of measurements
that can be performed together.
Examples of measurement scenarios include multipartite Bell-type scenarios
familiar from discussions of nonlocality, Kochen–Specker configurations, measure-
ment scenarios associated with qudit stabiliser quantum mechanics, and more. For
example, the Bell scenario from Sect. 1.2.1, where two experimenters, Alice and
Bob, can each choose between performing one of two different measurements, a
or a  for Alice and b or b for Bob, obtaining one of two possible outcomes, is
represented as follows:

X = {a, a  , b, b} O = {0, 1}
M = {{a, b}, {a, b }, {a , b}, {a  , b }}.
 

Given this description of the experimental setup, then either performing repeated
runs of such experiments with varying choices of measurement context and
recording the frequencies of the various outcome events, or calculating theoretical
predictions for the probabilities of these outcomes, results in a probability table like
that in Sect. 1.2.1.
Such data is formalised as an empirical model for the given measurement
scenario X, M, O. For each valid choice of measurement context, it specifies the
probabilities of obtaining the corresponding joint outcomes. That is, it is a family
{eC }C∈M where each eC is a probability distribution on the set O C of functions
assigning an outcome in O to each measurement in C (the rows of the probability
table).
We require that the marginals of these distributions agree whenever contexts
overlap, i.e.

∀C, C  ∈ M. eC |C∩C  = eC  |C∩C  ,

where the notation eC |U with U ⊆ C stands for marginalisation of probability

(to ‘forget’ the outcomes of some measurements): for t ∈ O U ,
distributions 
eC |U (t) := s∈O C ,s|U =t eC (t). The requirement of compatibility of marginals is
a generalisation of the usual no-signalling condition, and is satisfied in particular
by all empirical models arising from quantum predictions (Abramsky and Branden-
burger 2011).
An empirical model is said to be non-contextual if this family of distributions
can be obtained as the marginals of a single probability distribution on global
assignments of outcomes to all measurements, i.e. a distribution d on O X
1 Classical Logic, Classical Probability, and Quantum Mechanics 9

(where O X acts as a canonical set of deterministic hidden variables) such that

∀C ∈ M. d|C = eC . Otherwise, it is said to be contextual. Equivalently Abramsky
and Brandenburger (2011), contextual empirical models are those which have no
realisation by factorisable hidden variable models; thus for Bell-type measurement
scenarios contextuality specialises to the usual notion of nonlocality.
Noncontextuality characterizes classical behaviours. One way to understand this
is that it reflects a situation in which the physical system being measured exists
at all times in a definite state assigning outcome values to all properties that can be
measured. Probabilistic behaviour may still arise, but only via stochastic mixtures or
distributions on these global assignments. This may reflect an averaged or aggregate
behaviour, or an epistemic limitation on our knowledge of the underlying global
assignment.

1.3.2 The Non-contextual Polytope

Suppose we are given a measurement scenario X, M, O. Each global assignment
t ∈ O X induces a deterministic empirical model δ t :

1, t|C = s
δ tC (s) =
0 otherwise.

Note that t|C is the function t restricted to C, which is a subset of its domain X.
We have the following result from (Abramsky and Brandenburger 2011, Theorem
8.1):
Theorem 2 An empirical model {e C } is non-contextual if and only if it can be
written as a convex combination j ∈J μj δ tj where tj ∈ O X for each j ∈ J .
This means that for each C ∈ M,
 t
eC = μj δ Cj .
j


Given a scenario X, M, O, define m := C∈M |O C | = | {C, s | C ∈ M,
s ∈ O C | to be the number of joint outcomes as we range over contexts. For
example, in the case of the Bell table there are four contexts, each with four possible
outcomes, so m = 16. We can regard an empirical model {eC } over X, M, O as a
real vector ve ∈ Rm , with ve [C, s] = eC (s).
We define the non-contextual polytope for the scenario X, M, O to be the
convex hull of the set of deterministic models δ t , where t ranges over O X . By the
preceding theorem, this is exactly the set of non-contextual models.
Thus we have captured the question as to whether an empirical model is
contextual in terms of membership of a polytope. The facet inequalities for this
family of polytopes, as we range over measurement scenarios, give a general notion
of Bell inequalities.
10 S. Abramsky

As explained in Abramsky and Hardy (2012), a complete finite set of rational

facet inequalities for the contextual polytope over a measurement scenario can be
computed using Fourier-Motzkin elimination. This procedure is doubly-exponential
in the worst case, but standard optimizations reduce this to a single exponen-
tial. Despite this high complexity, Fourier-Motzkin elimination is widely used
in computer-assisted verification and polyhedral computation (Strichman 2002;
Christof et al. 1997).

1.3.3 Completeness of Logical Bell Inequalities

Suppose we are given a scenario X, M, O. A rational inequality is given by

a rational vector r and a rational number r. An empirical model v satisfies this
inequality if r · v ≤ r. Two inequalities are equivalent if they are satisfied by the
same empirical models.
Theorem 3 A rational inequality is satisfied by all non-contextual empirical
models over X, M, O if and only if it is equivalent to a logical Bell inequality
of the form (1.2).
For the proof, see Abramsky and Hardy (2012). Combining this result with the
previous observations, we obtain:
Theorem 4 The polytope of non-contextual empirical models over any scenario
X, M, O is determined by a finite set of logical Bell inequalities. Moreover, these
inequalities can be obtained effectively from the scenario. Thus an empirical model
over any scenario is contextual if and only if it violates one of finitely many logical
Bell inequalities.
It is worth reflecting on the conceptual import of these results. They are saying
that the scope for non-classical correlations available to quantum mechanics, or
any other physical theory, arises entirely from families of events which are locally
consistent, but globally inconsistent. Thus non-classical probabilistic behaviour
rests on this kind of logical structure.

1.4 The Contextual Fraction

We now turn to computational aspects. Fix a measurement scenario X, M, O.

 n := |O | be Cthe  number of global assignments g, and m :=
Let X

| C, s | C ∈ M, s ∈ O | be the number of local assignments ranging over

contexts. The incidence matrix Abramsky and Brandenburger (2011) M is an
m × n (0, 1)-matrix that records the restriction relation between global and local
assignments:
1 Classical Logic, Classical Probability, and Quantum Mechanics 11

1 if g|C = s;
M[C, s, g] :=
0 otherwise.

As already explained, an empirical model e can be represented as a vector ve ∈ Rm ,

with the component ve [C, s] recording the probability given by the model to
the assignment s at the measurement context C, eC (s). This vector is a flattened
version of the table used to represent the empirical model. The columns of the
incidence matrix, M[−, g], are the vectors corresponding to the (non-contextual)
deterministic models δ g obtained from global assignments g ∈ O X . Recall that
every non-contextual model can be written as a convex combination of these. A
probability distribution on global assignments can be represented as a vector d ∈ Rn
with non-negative components, and then the corresponding non-contextual model is
represented by the vector M d. So a model e is non-contextual if and only if there
exists d ∈ Rn such that:

M d = ve and d ≥ 0.

It is also natural to consider a relaxed version of this question, which leads us to

the contextual fraction.
Given two empirical models e and e on the same measurement scenario and
λ ∈ [0, 1], we define the empirical model λe + (1 − λ)e by taking the convex
sum of probability distributions at each context. Compatibility is preserved by this
convex sum, hence it yields a well-defined empirical model.
A natural question to ask is: what fraction of a given empirical model e admits
a non-contextual explanation? This approach enables a refinement of the binary
notion of contextuality vs non-contextuality into a quantitative grading. Instead of
asking for a probability distribution on global assignments that marginalises to the
empirical distributions at each context, we ask only for a subprobability distribution4
b on global assignments O X that marginalises at each context to a subdistribution of
the empirical data, thus explaining a fraction of the events, i.e. ∀C ∈ M. b|C ≤ eC .
Equivalently, we ask for a convex decomposition

e = λeN C + (1 − λ)e (1.3)

where eN C is a non-contextual model and e is another (no-signalling) empirical

model. The maximum weight of such a global subprobability distribution, or the
maximum value of λ in such a decomposition,5 is called the non-contextual fraction
of e, by analogy with the local fraction previously introduced for models on Bell-

4A subprobabilitydistribution on a set S is a map b : S −→ R≥0 with finite support and w(b) ≤ 1,

where w(b) := s∈S b(s) is called its weight. The set of subprobability distributions on S is
ordered pointwise: b is a subdistribution of b (written b ≤ b) whenever ∀s ∈ S. b (s) ≤ b(s).
5 Note that such a maximum exists, i.e. that the supremum is attained. This follows from the

Heine–Borel and extreme value theorems since the set of such λ is bounded and closed.
12 S. Abramsky

type scenarios Elitzur et al. (1992). We denote it by NCF(e), and the contextual
fraction by CF(e) := 1 − NCF(e).
The notion of contextual fraction in general scenarios was introduced in Abram-
sky and Brandenburger (2011), where it was proved that a model is non-contextual
if and only if its non-contextual fraction is 1.
A global subprobability distribution is represented by a vector b ∈ Rn with non-
negative components, its weight being given by the dot product 1 · b, where 1 ∈ Rn
is the vector whose n components are each 1. The following LP thus calculates the
non-contextual fraction of an empirical model e:

Find b ∈ Rn
maximising 1 · b
(1.4)
subjectto M b ≤ ve
and b ≥ 0 .

An inequality for a scenario X, M, O is given by a vector a ∈ Rm of real

coefficients indexed by local assignments C, s, and a bound R. For a model e, the
inequality reads a · ve ≤ R, where

a · ve = a[C, s] eC (s).
C∈M,s∈O C

Without loss of generality, we can take R to be non-negative (in fact, even R = 0)

as any inequality is equivalent to one of this form. We call it a Bell inequality if it is
satisfied by every non-contextual model. If, moreover, it is saturated by some non-
contextual model, the Bell inequality is said to be tight. A Bell inequality establishes
a bound for the value of a · ve amongst non-contextual models e. For more general
models, this quantity is limited only by the algebraic bound6


a := max a[C, s] | s ∈ O C .
C∈M

The violation of a Bell inequality a, R by a model e is max{0, a·ve −R}. However,
it is useful to normalise this value by the maximum possible violation in order to give
a better idea of the extent to which the model violates the inequality. The normalised
violation of the Bell inequality by the model e is

max{0, a · ve − R}
.
a − R

6 We will consider only inequalities satisfying R < a, which excludes inequalities trivially

satisfied by all models, and avoids cluttering the presentation with special caveats about division
by 0.
1 Classical Logic, Classical Probability, and Quantum Mechanics 13

Theorem 5 Let e be an empirical model.

1. The normalised violation by e of any Bell inequality is at most CF(e);
2. if CF(e) > 0, this bound is attained, i.e. there exists a Bell inequality whose
normalised violation by e is CF(e);
3. moreover, for any decomposition of the form e = NCF(e)eN C + CF(e)eSC , this
Bell inequality is tight at the non-contextual model eN C (provided NCF(e) > 0)
and maximally violated at the strongly contextual model eSC .
The proof of this result is based on the Strong Duality theorem of linear
programming (Dantzig and Thapa 2003). This provides an LP method of calculating
a witnessing Bell inequality for any empirical model e. For details, see the
supplemental material in Abramsky et al. (2017).

1.5 Remarks on Complexity

We now return to the issue of the complexity of deciding whether a given empirical
model is contextual.
It will be useful to consider the class of (n, k, 2) Bell scenarios. In these scenarios
there are n agents, each of whom has a choice of k measurement settings, and all
measurements have 2 possible outcomes. This gives rise to a measurement scenario
X, M, O where |X| = nk. Each C ∈ M consists of n measurements, one chosen
by each agent. Thus |M| = k n . For each context C, there are 2n possible outcomes.
An empirical model for an (n, k, 2) Bell scenario is thus given by a vector of
k n 2n probabilities. Thus the size of instances is exponential in the natural parameter
n. This is the real obstacle to tractable computation as we increase the number of
agents.
Given an empirical model ve as an instance, we can use the linear program (1.4)
given in the previous section to determine if it is contextual. The size of the linear
program is determined by the incidence matrix, which has dimensions p × q, where
p is the dimension of ve , and q = 2nk .
If we treat n as the complexity parameter, and keep k fixed, then q = O(s k ),
where s is the size of the instance. Thus the linear program has size polynomial
in the size of the instance, and membership of the non-contextual polytope can be
decided in time polynomial in the size of the instance.
There is an interesting contrast with Pitowsky’s results on the NP-completeness
of deciding membership in the correlation polytope of all binary conjunctions of
basic events. In Pitowsky’s case, the size of instances for these special forms of
correlation polytope is polynomial in the natural parameter, which is the number
of basic events. If we consider the correlation polytopes which would correspond
directly to empirical models, the same argument as given above would apply: the
instances would have exponential size in the natural parameter, while membership
in the polytope could be decided by linear programming in time polynomial in the
instance size.
14 S. Abramsky

We can also consider the situation where we fix n, and treat k as the complexity
parameter. In this case, note that the size of instances are polynomial in k, while the
size of the incidence matrix is exponential in k. Thus linear programming does not
help. In fact, this seems to be the case that Pitowsky primarily had in mind. With
n = 2, the restriction to binary conjunctions makes good sense, and all the examples
he discusses are of this kind.
We also mention the results obtained in Abramsky and Hardy (2012); Abramsky
et al. (2013); Abramsky (2017); Abramsky et al. (2017), which study the analogous
problem with respect to possibilistic contextuality, that is, whether the supports
of the probability distributions in the empirical model can be obtained from a
set of global assignments. In that case, the complexity of deciding contextuality
is shown to be NP-complete in the complexity parameter k in general; a precise
delineation is given of the tractability boundary in terms of the values of the
parameters. Moreover, as Rui Soares Barbosa has pointed out,7 if we take n as
the complexity parameter, there is a simple algorithm for detecting possibilistic
contextuality which is polynomial in the size of the instance. Thus the complexity
of detecting possibilistic contextuality runs completely in parallel with the proba-
bilistic case.

1.6 The “Edge of Logical Contradiction” vs. the “Boundary

of Paradox”

We give a final quotation from (Pitowsky 1994, p. 113):

A violation of Boole’s conditions of possible experience cannot be encountered when all
the frequencies concerned have been measured on a single sample. Such a violation simply
entails a logical contradiction; ‘observing’ it would be like ‘observing’ a round square. We
expect Boole’s conditions to hold even when the frequencies are measured on distinct large
random samples. But they are systematically violated, and there is no easy way out (see
below). We thus live ‘on the edge of a logical contradiction’. An interpretation of quantum
mechanics, an attempt to answer the WHY question, is thus an effort to save logic.

In my view, this states the extent of the challenge posed to classical logic by quantum
mechanics too strongly. As we have discussed, the observational data predicted
by quantum mechanics and confirmed by actual experiments consists of families
of probability distributions, each defined on different sample spaces, correspond-
ing to the different contexts. Since the contexts overlap, there are relationships
between the sample spaces, which are reflected in coherent relationships between
the distributions, in the form of consistent marginals. But there is no “global”
distribution, defined on a sample space containing all the observable quantities,
which accounts for all the empirically observable data. This does pose a challenge
to the understanding of quantum mechanics as a physical theory, since it implies

7 Personal communication.
1 Classical Logic, Classical Probability, and Quantum Mechanics 15

that we cannot ascribe definite values to the physical quantities being measured,
independent of whether or in what context they are measured. It does not, however,
challenge classical logic and probability, which can be used to describe exactly this
situation.
For this reason, I prefer to speak of contextuality as living “on the boundary of
paradox” Abramsky (2017), as a signature of non-classicality, and one for which
there is increasing evidence that it plays a fundamental rôle in quantum advantage
in information-processing tasks (Anders and Browne 2009; Raussendorf 2013;
Howard et al. 2014). So this boundary seems likely to prove a fruitful place to be.
But we never actually cross the boundary, for exactly the reasons vividly expressed
by Pitowsky in the opening two sentences of the above quotation.
There is much more to be said about the connections between logic and
contextuality. In particular:
• There are notions of possibilistic and strong contextuality, and of All-versus-
Nothing contextuality, which give a hierarchy of strengths of contextuality, and
which can be described in purely logical terms, without reference to probabilities
(Abramsky and Brandenburger 2011).
• These “possibilistic” forms of contextuality can be connected with logical para-
doxes in the traditional sense. For example, the set of contradictory propositions
used in Sect. 1.2.1 to derive Bell’s theorem form a “Liar cycle” (Abramsky et al.
2015).
• There is also a topological perspective on these ideas. The four propositions from
Sect. 1.2.1 form a discrete version of the Möbius strip. Sheaf cohomology can be
used to detect contextuality (Abramsky et al. 2015).
• The logical perspective on contextuality leads to the recognition that the same
structures arise in many non-quantum areas, including databases (Abramsky
2013), constraint satisfaction (Abramsky et al. 2013), and generic inference
(Abramsky and Carù 2019).
• In Abramsky et al. (2017), an inequality of the form pF ≥ NCF(e)d is derived
for several different settings involving information-processing tasks. Here pF
is the failure probability, NCF(e) is the non-contextual fraction of an empirical
model viewed as a resource, and d is a parameter measuring the “difficulty”
of the task. Thus the inequality shows the necessity of increasing the amount
of contextuality in the resource in order to increase the success probability. In
several cases, including certain games, communication complexity, and shallow
circuits, the parameter d is n−K
n , where K is the K-consistency we encountered
in formulating logical Bell inequalities (1.2).
We refer to the reader to the papers cited above for additional information on
these topics.
16 S. Abramsky

1.7 Concluding Remarks

Itamar Pitowsky’s work on quantum foundations combines lucid analysis, con-

ceptual insights and mathematically sophisticated and elegant results. We have
discussed some recent and ongoing work, and related it to his contributions, which
remain a continuing source of inspiration.

Acknowledgements My thanks to Meir Hemmo and Orly Shenker for giving me the opportunity
to contribute to this volume in honour of Itamar Pitowsky. I had the pleasure of meeting Itamar on
several occasions when he visited Oxford.
I would also like to thank my collaborators in the work I have described in this paper: Adam
Brandenburger, Rui Soares Barbosa, Shane Mansfield, Kohei Kishida, Ray Lal, Giovanni Carù,
Lucien Hardy, Phokion Kolaitis and Georg Gottlob. My thanks also to Ehtibar Dzhafarov, whose
Contextuality-by-Default theory (Dzhafarov et al. 2015) has much in common with my own
approach, for our ongoing discussions.

References

Abramsky, S. (2013). Relational databases and Bell’s theorem. In In search of elegance in the
theory and practice of computation (pp. 13–35). Berlin: Springer.
Abramsky, S. (2017). Contextuality: At the borders of paradox. In E. Landry (Ed.), Categories for
the working philosopher. Oxford: Oxford University Press.
Abramsky, S., & Brandenburger, A. (2011). The sheaf-theoretic structure of non-locality and
contextuality. New Journal of Physics, 13(11), 113036.
Abramsky, S., & Carù, G. (2019). Non-locality, contextuality and valuation algebras: a general
theory of disagreement. Philosophical Transactions of the Royal Society A, 377(2157),
20190036.
Abramsky, S., & Hardy, L. (2012). Logical Bell inequalities. Physical Review A, 85(6), 062114.
Abramsky, S., Gottlob, G., & Kolaitis, P. (2013). Robust constraint satisfaction and local hidden
variables in quantum mechanics. In Twenty-Third International Joint Conference on Artificial
Intelligence.
Abramsky, S., Barbosa, R. S., Kishida, K., Lal, R., & Mansfield, S. (2015). Contextuality,
cohomology and paradox. In S. Kreutzer (Ed.), 24th EACSL Annual Conference on Computer
Science Logic (CSL 2015), volume 41 of Leibniz International Proceedings in Informatics
(LIPIcs) (pp. 211–228). Dagstuhl, Schloss Dagstuhl–Leibniz-Zentrum für Informatik.
Abramsky, S., Barbosa, R. S., & Mansfield, S. (2017). Contextual fraction as a measure of
contextuality. Physical Review Letters, 119(5), 050504.
Anders, J., & Browne, D. E. (2009). Computational power of correlations. Physical Review Letters,
102(5), 050502.
Bell, J. S. (1964). On the Einstein-Podolsky-Rosen paradox. Physics, 1(3), 195–200.
Boole, G. (1862). On the theory of probabilities. Philosophical Transactions of the Royal Society
of London, 152, 225–252.
Christof, T., Löbel, A., & Stoer, M. (1997). PORTA-POlyhedron Representation Transformation
Algorithm. Publicly available via ftp://ftp.zib.de/pub/Packages/mathprog/polyth/porta.
Dantzig, G. B., & Thapa, M. N. (2003). Linear programming 2: Theory and extenstions (Springer
series in operations research and financial engineering). New York: Springer.
Dzhafarov, E. N., Kujala, J. V., & Cervantes, V. H. (2015). Contextuality-by-default: A brief
overview of ideas, concepts, and terminology. In International Symposium on Quantum
Interaction (pp. 12–23). Heidelberg: Springer.
1 Classical Logic, Classical Probability, and Quantum Mechanics 17

Elitzur, A. C., Popescu, S., & Rohrlich, D. (1992). Quantum nonlocality for each pair in an
ensemble. Physics Letters A, 162(1), 25–28.
Ghirardi, G. C., Rimini, A., & Weber, T. (1980). A general argument against superluminal
transmission through the quantum mechanical measurement process. Lettere Al Nuovo Cimento
(1971–1985), 27(10), 293–298.
Howard, M., Wallman, J., Veitch, V., & Emerson, J. (2014). Contextuality supplies the ‘magic’ for
quantum computation. Nature, 510(7505), 351–355.
Jordan, T. F. (1983). Quantum correlations do not transmit signals. Physics Letters A, 94(6–7), 264.
Pitowsky, I. (1989). Quantum probability, quantum logic (Lecture notes in physics, Vol. 321).
Berlin: Springer.
Pitowsky, I. (1991). Correlation polytopes: Their geometry and complexity. Mathematical Pro-
gramming, 50(1–3), 395–414.
Pitowsky, I. (1994). George Boole’s “Conditions of possible experience” and the quantum puzzle.
The British Journal for the Philosophy of Science, 45(1), 95–125.
Raussendorf, R. (2013). Contextuality in measurement-based quantum computation. Physical
Review A, 88(2), 022322.
Strichman, O. (2002). On solving Presburger and linear arithmetic with SAT. In Formal methods
in computer-aided design (pp. 160–170). Heidelberg: Springer.
Weisstein, E. W. (2019). Bonferroni inequalities. http://mathworld.wolfram.com/
BonferroniInequalities.html. MathWorld–A Wolfram Web Resource.
Chapter 2
Why Scientific Realists Should Reject
the Second Dogma of Quantum
Mechanics

Valia Allori

Abstract The information-theoretic approach to quantum mechanics, proposed

by Bub and Pitowsky, is a realist approach to quantum theory which rejects the
“two dogmas” of quantum mechanics: in this theory measurement results are not
analysed in terms of something more fundamental, and the quantum state does
not represent physical entities. Bub and Pitowsky’s approach has been criticized
because of their argument that kinematic explanations are more satisfactory than
dynamical ones. Moreover, some have discussed the difficulties the information-
theoretic interpretation faces in making sense of the quantum state as epistemic.
The aim of this paper is twofold. First, I argue that a realist should reject the
second dogma without relying on the alleged explanatory superiority of kinematical
explanation over dynamical ones, thereby providing Bub and Pitowsky with a way
to avoid the first set of objections to their view. Then I propose a functionalist
account of the wavefunction as a non-material entity which does not fall prey of
the objections to the epistemic account or the other non-material accounts such as
the nomological view, and therefore I supply the proponents of the information-
theoretic interpretation with a new tool to overcome the second set of criticisms.

Keywords Two dogmas · Quantum information · Wave-function ontology ·

Primitive ontology · Functionalism · Epistemic interpretation of the
wave-function · Nomological interpretation of the wave-function

2.1 Introduction

Bub and Pitowsky (2010) have proposed the so-called information-theoretic (IT)
approach to quantum mechanics. It is a realist approach to quantum theory which
rejects what Bub and Pitowsky call the ‘two dogmas’ of quantum mechanics.

V. Allori ()
Department of Philosophy, Northern Illinois University, Dekalb, IL, USA
e-mail: vallori@niu.edu

© Springer Nature Switzerland AG 2020 19

M. Hemmo, O. Shenker (eds.), Quantum, Probability, Logic,
Jerusalem Studies in Philosophy and History of Science,
https://doi.org/10.1007/978-3-030-34316-3_2
20 V. Allori

These dogmas are as follows: (D1) measurement results should not be taken
as unanalysable primitives; and (D2) the wavefunction represents some physical
reality. In contrast, in the IT approach measurements are not analysed in terms of
something more fundamental, and the wavefunction does not represent physical
entities. Bub and Pitowsky reject the first dogma arguing that their approach is
purely kinematic, and that dynamical explanations (used to account for experimental
outcomes) add no further insight. Their approach has been criticized by Brown
and Timpson (2006) and by Timpson (2010), who argue that kinematic theories
should not be a model for realist quantum theories. The second dogma is rejected
as a consequence of the rejection of the first. Bub and Pitowsky assume that their
rejection of the wavefunction as representing physical entities commits them to an
epistemic view in which the wavefunction reflects is a reflection of our ignorance,
and as such they have been criticized by some authors, such as Leifer (2014) and
Gao (2017).
Little has been done to investigate whether it is possible to reject the second
dogma without appealing to the alleged superiority of dynamical explanations, and
whether one could think of the wavefunction as not representing matter without
falling prey to the objections to the epistemic view (or to the so-called nomological
view). The aim of this paper is to explore such issues. Therefore, first I argue that
Bub and Pitowsky, as well as scientific realists in general, should reject the second
dogma to start with, independently of whether they decide to accept or deny the
first dogma. This argument supports the IT account to motivate their view, but it
also stands alone as an argument to reject the wavefunction as a material entity
in general. Then, in connection with this, I propose a functional account of the
wavefunction which avoids the objections of the alternative views. This nicely fits
in the IT interpretation by dissolving the objections against it which relies on the
wavefunction being epistemic, but it is also helpful in answering general questions
about the nature of the wavefunction.
Here’s the scheme of the paper. In Sect. 2.2 I present the IT approach of Bub
and Pitowsky. Then in Sect. 2.3 I discuss the distinction between dynamical and
kinematical explanations, as well as the principle/constructive theory distinction,
and how they were used by Bub and Pitwosky to motivate their view. Section 2.4
concludes this part of the paper by presenting the objections to the view connected
with the explanatory superiority of kinematic theories. In Sect. 2.5 I move to the
most prominent realist view which accepts dogma 2, namely wavefunction realism.
Then in Sect. 2.6 I provide alternative reasons why a realist should reject the second
dogma by introducing the primitive ontology approach. I connect the rejection of
the first dogma to scientific realism in Sect. 2.7, showing how a realist may or may
not consistently reject dogma 1. Then I move to the second set of objections to
the IT account based on the epistemic wavefunction. First, in Sect. 2.8 I review the
objections to the ontic accounts, while in Sect. 2.9 I discuss the ones to the epistemic
account. Finally in Sect. 2.10 I propose an account of how the wavefunction based
on functionalism, and I argue that it is better than the alternatives. In the last section
I summarize my conclusions.
2 Why Scientific Realists Should Reject the Second Dogma of Quantum Mechanics 21

2.2 The IT Account, the Two Dogmas, and the Measurement

Problem

The IT approach to quantum theory is a realist interpretation of the quantum

formalism according to which quantum mechanics is fundamentally about measure-
ment outcomes, whose statistics are recovered by the constraints that Hilbert space
structure imposes on the quantum dynamics, and in which the wavefunction has the
role of describing an agent’s preferences (or expectations or degrees of belief) about
measurement outcomes. The approach comes at the end of two lines of research first
developed independently and then together.1
There are different components to the view, the first of which is given by
the CBH theorem (Clifton et al. 2003), which offers an understanding of the
quantum formalism in terms of simple physical principles,2 which can be suitably
translated into information-theoretic constraints on the probability of experimental
outcomes. In light of this, Bub (2004, 2005) argues that quantum mechanics is
about ‘information,’ taken as primitive. In this way, there is a sense in which
the CBH constraints make quantum mechanics a kinematic theory (see Sect.
2.3): experimental outcomes are fundamental, and the theory poses constraints on
what these outcomes are supposed to be. Bub also claims that further dynamical
explanations are not necessary and therefore quantum mechanics does not need any
additional fundamental ontology, even if is possible to provide one.3
The second ingredient of the IT account is given by Pitowsky’s idea that quantum
mechanics provides us with a new probability theory (Pitowsky 2007). Like classical
probability theory consists of a space of possible events and a measure over it,
here the space of possible events is the lattice of the closed subspaces of Hilbert
space. Thus, the event – space defines the possible experiment outcomes (which
are the fundamental elements of the theory) and the wavefunction encodes the
experimenter’s beliefs about them. According to Pitowsky, the advantage of this
approach is that it can ‘dissolve’ the measurement problem. First he argues that
there are two measurement problems:
• The “big” measurement problem: the problem of providing a dynamical expla-
nation of why particular experiments have the outcomes they do; and

1 See,respectively Clifton et al. (2003), Bub (2004, 2005, 2007), Pitowsky (2007), Bub and
Pitowsky (2010).
2 Here they are: the impossibility of superluminal information transfer between two physical

systems by performing measurements on one of them, the impossibility of broadcasting the

information contained in an unknown physical state, and the impossibility of unconditionally
secure bit commitment.
3 He writes: “you can, if you like, tell a story along Bohmian, or similar, lines [ . . . ] but, given the

information-theoretic constraints, such a story can, in principle, have no excess empirical content
over quantum mechanics” (Bub 2005, p. 542).
22 V. Allori

• The “small” measurement problem: the problem of explaining how an effectively

classical probability space of macroscopic measurement outcomes arises from a
quantum measurement process.
And then he argues that the ‘big’ one, which is what is usually called the
measurement problem, is actually a pseudo-problem, given that in his approach one
cannot talk about one single experimental results, just only the statistics.
One immediate reaction to this interpretation is that it is not truly realist,
and therefore it is not in direct competition with the ‘traditional’ realist quantum
theories such as the pilot-wave theory (Bohm 1952), the spontaneous collapse
theory (Ghirardi et al. 1986) and Everettian mechanics (Everett 1957). However,
Bub and Pitowsky argue that there is a robust sense in which their intent is realist,
and that people who deny this implicitly accepts on two ‘dogmas,’ which they think
have little justification:
D1: measurement should never be included as an unanalyzable primitive in a
fundamental physical theory, and
D2: the wavefunction represents physical reality.
Bub and Pitowsky claim that it is usually thought that realists have to accept them to
solve the measurement problem. However, they say, this is true only if they have
in mind the ‘big’ problem, but in order to be realist one merely needs to solve
the ‘small’ problem. In fact, they point out, take the Everett interpretation: it is
considered a realist theory even if it only solves the ‘small’ problem. In Everett’s
theory, in fact, measurements have no definite outcomes and the experimental
statistics are obtained in terms of a many-worlds structure. This structure acts
classically due to decoherence plus the idea of considering probabilities as an
agent’s preferences (Wallace 2002). For Bub and Pitowsky, the difference between
their approach and Everett is that while Everettians accept the two dogmas, they
do not. However, they have in common that they still solve the ‘small’ measurement
problem, given that in the IT account measurements are taken to be primitive, which
is readily solved by the constraints imposed by the Hilbert space structure. This, they
argue, shows that one could be realist even if rejecting the two dogmas of quantum
mechanics. (In Sect. 2.5 I will provide another example of a realist quantum theory
which rejects dogma 2).
Moreover, building on what Bub previously did, Bub and Pitowsky also argue
that one should reject both dogmas. They do this by appealing to the distinction
between dynamical and kinematic explanation, connected with the notions of
principle and constructive theories which was originally introduced by Einstein
(1919). They argue that kinematic theories are more satisfactory that dynamical
ones, and that, given that the IT approach is the only realist interpretation that could
be thought as a kinematic theory, it is to be preferred to the alternatives. Since this
theory rejects the two dogmas, so should the realist. Let’s discuss this argument in
the next section.
2 Why Scientific Realists Should Reject the Second Dogma of Quantum Mechanics 23

2.3 The IT Approach as a Kinematic Theory

Bub and Pitowsky argue that the IT interpretation is a kinematic theory, and that
this kind of theories are to be preferred, as dynamical theories add no further
insight. Originally, the argument has been presented by Bub in terms of the
principle/constructive theories distinction, so let’s start with that. Principle theories
are formulated in terms principles, which are used as constraints on physically pos-
sible processes, as in thermodynamics (‘no perpetual motion machines’). Instead,
constructive theories involve the dynamical reduction of macroscopic objects in
terms of their microscopic constituents, as in the kinetic theory (which reduces the
behavior of gases to the motion of atoms).4 Einstein introduced this distinction when
discussing his 1905 theory of relativity, which he regarded as a principle theory, as
it was formulated in terms of the two principles of equivalence of inertial frames for
all physical laws, and constancy of the velocity of light (in vacuum for all inertial
frames). This theory explains relativistic effects (such as length contraction and
time dilation) as the physical phenomena compatible with the theory’s principles.
By contrast, since Lorentz’s theory (1909) derives the relativistic transformations
and the relativistic effects from the electromagnetic properties of the ether and its
interactions with matter, is a constructive theory.5
According to Bub and Pitowsky, Lorentz’s constructive theory came first, then
only later Einstein formulated his (principle theory of) special relativity: Minkowski
provided kinematic constraints which relativistic dynamics has to obey, and then
Einstein came up with an interpretation for special relativity. Bub and Pitowsky
also claim that we should take the fact that Einstein’s kinematic theory has been
preferred over Lorentz’s dynamical theory as evidence that such type of theories
are to be preferred in general. In addition, they argue that the fact that Lorentz’s
theory can constructively explain Lorentz invariance justified a realist interpretation
of special relativity as a principle theory. But after this initial use, the constructive
counterpart is no longer necessary. Similarly, Bub and Pitowsky argue that IT
is kinematic theory in that it provides constraints on the phenomena without

4 Here’s how Balashov and Janssen put it:“In a theory of principle [ . . . ] one explains the
phenomena by showing that they necessarily occur in a world in accordance with the postulates.
Whereas theories of principle are about the phenomena, constructive theories aim to get at the
underlying reality. In a constructive theory one proposes a (set of) model(s) for some part of
physical reality [ . . . ]. One explains the phenomena by showing that the theory provides a model
that gives an empirically adequate description of the salient features of reality” (Balashov and
Janssen 2003).
5 Again in the worlds of Balashov and Janssen: “Consider the phenomenon of length contraction.

Understood purely as a theory of principle, SR explains this phenomenon if it can be shown that
the phenomenon necessarily occurs in any world that is in accordance with the relativity postulate
and the light postulate. By its very nature such a theory-of-principle explanation will have nothing
to say about the reality behind the phenomenon. A constructive version of the theory, by contrast,
explains length contraction if the theory provides an empirically adequate model of the relevant
features of a world in accordance with the two postulates. Such constructive-theory explanations
do tell us how to conceive of the reality behind the phenomenon” (Balashov and Janssen 2003).
24 V. Allori

explaining them dynamically. That is, Hilbert space should be recognized as a

“the kinematic framework for the physics of an indeterministic universe, just as
Minkowski space-time provides the kinematic framework for the physics of a
non-Newtonian, relativistic universe” (Bub and Pitowsky 2010, emphasis in the
original text). Because of this, no other explanation for the experimental results is
necessary.6 More in detail, “the information-theoretic view of quantum probabilities
as ‘uniquely given from the start’ by the structure of Hilbert space as a kinematic
framework for an indeterministic physics is the proposal to interpret Hilbert space
as a constructive theory of information-theoretic structure or probabilistic structure”
(Bub and Pitowsky 2010).
As a consequence, we should reject the first dogma to obtain a kinematic theory,
not a dynamical one. According to Bub and Pitowsky, the ‘big’ measurement
problem is a dynamical problem, and as such a pseudo-problem when considering
kinematic theories like theirs: as relativistic effects, such as length contraction and
time dilation, are a problem for Newtonian mechanics but ceases to be such when
looking at the geometry of Minkowski space-time and thus at the kinematics, here
the ‘big’ problem dissolves when looking at the constraints that the Hilbert space
structure imposes on physical events, given by experimental outcomes.
The next step is to observe that by rejecting dogma 1, the wavefunction can only
be connected to experimenters’ degrees of belief, rather than representing something
real. Because of this, therefore, it is argued that we should also reject dogma 2.
In other words, the rejection of the first dogma allows to make quantum theory a
kinematic theory, which in turns dissolves the ‘big’ measurement problem, while
the role of the rejection of the second dogma is to find a place for the wavefunction
in the interpretation. That is, it allows to answer the question: if measurements are
primitive, what is the wavefunction?
Be that as it may, in the next section we will start discussing some criticism raised
against the IT account.

2.4 Objections to the Explanatory Superiority of Kinematic

Theories

The IT framework has been criticized for mainly two reasons: on the one hand,
objections have been raised against the alleged superiority of kinematic theories,
therefore undermining the rejection of dogma 1; and on the other hand, people
have pointed out the difficulties of considering the wavefunction as epistemic, as
a consequence of the rejection of dogma 2. In this section, we will review the first

6 “There is no deeper explanation for the quantum phenomena of interference and entanglement
than that provided by the structure of Hilbert space, just as there is no deeper explanation for the
relativistic phenomena of Lorentz contraction and time dilation than that provided by the structure
of Minkowski space-time” (Bub and Pitowsky 2010).
2 Why Scientific Realists Should Reject the Second Dogma of Quantum Mechanics 25

kind of criticisms, while in Sect. 2.9 we will discuss the ones based on the nature of
the wavefunction.
First of all, Timpson (2010) argues that the dogmas Bub and Pitowsky consider
are not dogmas at all, as they can be derived from more general principles like
realism. In fact, regarding dogma 1, a realist recognizes measurement results and
apparatuses as merely complicated physical systems without any special physical
status. Regarding dogma 2, we just proceed as we did in classical theories: there’s a
formalism, there’s an evolution equation, then one is realist about what the equation
is about, which, in the case of quantum mechanics, is the wavefunction. However, I
think Timpson misses the point. The dogmas are dogmas for the realist: the realist
is convinced that she has to accept them to solve the measurement problem. What
Bub and Pitowsky are arguing is that realist does not have to accept these dogmas:
Bub and Pitowsky in fact provide a realist alternative which does not rely on these
dogmas. Timpson is on target in that a realist will likely endorse dogma 1 (what’s
special about measurements?), but Bub and Pitowsky argue that that is not the only
option. Namely, one can recognize dogma 1 as being a dogma and therefore may
want to recognize the possibility of rejecting it. One advantage of doing so, as we
saw, is that in this way quantum theory becomes a kinematic theory, if one believes
in the explanatory superiority of these type of theories. This leads one to recognize
dogma 2 also as a dogma, and then to recognize the possibility of rejecting it as
well.
Moreover, let me notice that Timpson also points out that Bub and Pitowsky
provide us with no positive reason why one should be unhappy with such dogmas. I
will present in Sect. 2.9 reasons for being extremely suspicious of dogma 2. Indeed,
as I will argue later, one can restate the situation as follows: a realist, based on
the reasons Timpson points out, at first is likely to accept both dogmas. However,
upon further reflection, after realizing that by accepting dogma 2 one needs to face
extremely hard problems, the realist should reject it.
Brown and Timpson (2006) point out that the first dogma was not a dogma at
all for Einstein.7 In this way, they criticize the historical reconstruction of Bub
and Pitowsky for their argument for the superiority of kinematic theories. In the
view of Brown and Timpson, Einstein’s 1905 theory was seen by Einstein himself
as a temporary theory, to be discarded once a deeper theory would make itself

7 In fact, discussing his theory of special relativity he wrote: “One is struck that the theory [ . . . ]
introduces two kinds of physical things, i.e., (1) measuring rods and clocks, (2) all other things,
e.g., the electromagnetic field, the material point, etc. This, in a certain sense, is inconsistent;
strictly speaking measuring rods and clocks would have to be represented as solutions of the basic
equations (objects consisting of moving atomic configurations), not, as it were, as theoretically
self-sufficient entities. However, the procedure justifies itself because it was clear from the very
beginning that the postulates of the theory are not strong enough to deduce from them sufficiently
complete equations [ . . . ] in order to base upon such a foundation a theory of measuring rods and
clocks. [ . . . ] But one must not legalize the mentioned sin so far as to imagine that intervals are
physical entities of a special type, intrinsically different from other variables (‘reducing physics to
geometry’, etc.)” (Einstein 1949b).
26 V. Allori

available.8 Indeed, they point out, Einstein introduced the principle/constructive

distinction to express his own dissatisfaction of the theory at the time. According
to them, it was Einstein’s view that kinematic theories are typically employed when
dynamical theories are either unavailable or too difficult to build.9 In this way, they
argue that it is too much of a stretch to argue that Einstein would have encouraged
a kinematic theory interpretation of quantum mechanics, as Bub and Pitowsky
propose. However, even if I think that Brown and Timpson are correct, it is not clear
why this should be an issue for Bub and Pitowsky. In fact, the charge at this stage
is merely that Einstein did not consider dogma 1 as such, and that he did not like
kinematic explanations. And to this Bub and Pitowsky can simply reply that they
disagree with Einstein, and insist that there is value in kinematic theories, contrarily
to what Einstein himself thought.
Therefore, let’s move to the more challenging charge. Brown and Timpson
argue against the explanatory superiority of kinematic theories by pointing out that
only dynamical theories provide insight of the reality underlying the phenomena.10
That is, regularities and constraints over the possible experimental findings lack
explanatory power because we still do not know why these constraints obtain. A
reason is provided only by a dynamical, constructive theory, in which one is told
the mechanism, or the microscopic story, that gives rise to the observed behavior.
For instance, it has been argued that the economy of thermodynamic reasoning is
trumped by the insight that statistical mechanics provides (Bell 1976; Brown and
Timpson 2006).
Whether dynamical theories are actually more explanatory than kinematic
theories remains a controversial issue, and certainly more should be written in this
regard.11 Nonetheless, even if one were to accept that in some contexts (such as
in special relativity or thermodynamics) that dynamical theories are better, Bub
(2004) has argued that a kinematic quantum theory is superior to any dynamical
account. In fact, while Einstein’s theory of Brownian motion, which led to the
acceptance of the atomic theory of matter as more than a useful fiction, showed
the empirical limits of thermodynamics and thus the superiority of the dynamical
theory, no such advantage exists in the quantum domain. Bub in fact argues that such

8 “The methodology of Einstein’s 1905 theory represents a victory of pragmatism over explanatory
depth; and that its adoption only made sense in the context of the chaotic state of physics at the
start of the 20th century” (Brown and Timpson 2006).
9 For according to Einstein, “when we say we have succeeded in understanding a group of natural

processes, we invariably mean that a constructive theory has been found which covers the processes
in question” (Brown and Timpson 2006).
10 See also Brown (2005), and Brown and Pooley (2004).
11 See Flores (1999), who argues that principle theories, in setting out the kinematic structure, are

more properly viewed as explaining by unification, in the Friedman (1974)/Kitcher (1989) sense.
Moreover, see Felline (2011) for an argument against Brown’s view in the framework of special
relativity. See also van Camp (2011) who argues that it is not the kinematic feature of Einsteinian
relativity that makes it preferred to the Lorenzian dynamics but the constructive component, in
virtue of which the theory possesses a conceptual structure.
2 Why Scientific Realists Should Reject the Second Dogma of Quantum Mechanics 27

an advantage is given only when a dynamical theory can provide some ‘excess’
in empirical content. Since dynamical constructive quantum theories (such as the
pilot-wave theory) are empirically equivalent to a kinematic quantum theory, then
(unlike kinetic theory) no dynamical quantum theory can ever provide more insight
than a kinematic one. Critics however could point out immediately that there are
dynamical quantum theories such as the spontaneous collapse theory, which are not
empirically equivalent to quantum mechanics. Moreover, it is not obvious at all that
the only reason to prefer dynamical theories is because they can point to new data.
Therefore, it seems safe to conclude that an important step in Bub and Pitowsky’s
argument for their view rests on controversial issues connected with whether a
kinematic explanation is better than a dynamical one. Since the issue has not
been settled, I think it constitutes a serious shortcoming of the IT interpretation
as it stands now, and it would be nice if instead it could be grounded on a less
controversial basis. Indeed, this is what I aim to do in Sect. 2.6. Before doing this,
however, I will present in the next section the most popular realist approach which
accepts the second dogma.

2.5 Accepting the Second Dogma: Wavefunction Realism

Part of Bub and Pitowsky’s reasoning is based on their willingness to dissolve the
‘big’ measurement problem. So let’s take a step back and reconsider this problem
again. As is widely known, quantum theory has always been regarded as puzzling, at
best. As recalled by Timpson (2010), within classical mechanics the scientific realist
can read out the ontology from the formalism of the theory: the theory is about the
temporal evolution of a point in three-dimensional space, and therefore the world
is made of point-like particles. Similarly, upon opening books on quantum theory,
a mathematical entity and an evolution equation for it immediately stand out: the
wavefunction and the Schrödinger equation. So, the natural thought is to take this
object as describing the fundamental ontology. Even if we have just seen how it can
be readily justified, as we have seen, Bub and Pitowsky think it is a dogma, in the
sense that it is something that a realist has the option of rejecting. Here I want to
reinforce the argument that this is indeed a dogma, and that this dogma should be
rejected.
As emphasized by Schrödinger when discussing his cat paradox to present
the measurement problem (1935b), assuming the second dogma makes the theory
empirically inadequate. In fact, superposition states, namely states in which ‘stuff’
is both here and there at the same time, are mathematically possible solutions of
the Schrödinger equation. This feature is typical of waves, as the wavefunction is in
virtue of obeying Schrödinger’s equation (which is a wave equation). However, if
the wavefunction provides a complete description of the world, superpositions may
exist for everything, including cats in boxes being alive and dead at the same time.
The problem is that we never observe such macroscopic superpositions, and this
is, schematically, the measurement problem: if we open the box and we measure
28 V. Allori

the state of the cat, we find the cat either dead or alive. The realist, keeping the
second dogma true, traditionally identified their next task in finding theories to
solve this problem. That is, the realist concluded that, in Bell’s words, “either
the wavefunction, as given by the Schrödinger equation, is not everything, or it
is not right” (Bell 1987, 201). Maudlin (1995) has further elaborated the available
options by stating that the following three claims are mutually inconsistent: (A)
the wavefunction of a system is complete; (B) the wavefunction always evolves
to the Schrödinger equation; and (C) measurements have determinate outcomes.
Therefore, the most popular realist quantum theories are traditionally seen as
follows: the pilot-wave theory (Bohm 1952) rejects A, the spontaneous collapse
theory (Ghirardi et al. 1986) denies B, while the many-world theory (Everett 1957)
denies C.
Notice that it is surprising that even if dogma 2 has been challenged or resisted
in the 1920s by de Broglie, Heisenberg, Einstein and Lorentz (see next section),12
the physicists proposing the realist quantum theories mentioned above did not do
so. Indeed, the view which accepts the second dogma has been the accepted view
among the realists for a very long time. In this sense it was a dogma, as Bub and
Pitowsky rightly call it, until it was recently recognized to be an assumption and the
name of wavefunction realism was given to the view which accepts dogma 2.
Even if it is not often discussed, most wavefunction realists acknowledge that the
measurement problem leads to another problem, the so-called configuration space
problem. The wavefunction by construction is an object that lives on configuration
space. This is classically defined as the space of the configurations of all particles,
which, by definition, is a space with a very large number of dimensions. So, given
that for dogma 2 the ontology is given by the wavefunction, if physical space is
the space in which the ontology lives, then physical space is configuration space.
Accordingly, material objects are represented by a field in configuration space. That
is, the arena in which physical phenomena take place seems to be fundamentally
three-dimensional, but fundamentally it is not; and physical objects seem to be
three-dimensional, while fundamentally they are not. Thus, the challenge for the
wavefunction realist is to account for the fact that the world appears so different
from what it actually is.13 According to this view, the realist theories mentioned
above are all taken to be theories about the wavefunction, aside from the pilot-wave
theory which had both particles and waves. In contrast, since in all the other theories
everything is ‘made of’ wavefunctions, particles and particle-like behaviors are best
seen as emergent, one way or another.14
Be that as it may, the configuration space problem, after it was ignored since the
1920s, started to be recognized and discussed only in the 1990s, and at the moment it

12 See Bacciagaluppi and Valentini (2009) for an interesting discussion of the various positions
about this issue and others at the 1927 Solvay Congress.
13 See most notably Albert (1996, 2013, 2015), Lewis (2004, 2005, 2006, 2013), Ney (2012, 2013,

2015, 2017, forthcoming), North (2013).

14 See Albert (2015) and Ney (2017, forthcoming) for two different approaches to this.
2 Why Scientific Realists Should Reject the Second Dogma of Quantum Mechanics 29

is acknowledged that it has no universally accepted solution.15 Regardless, strangely

enough, most people keep identifying the measurement problem (who has been
solved) and not the configuration space problem (which has not) as the problem for
the realist.16 In the next section I will present a proposal to eliminate this problem
once for all.

2.6 Why Realists Should Reject the Second Dogma

In the last two decades or so, some authors have started to express resistance to the
characterization of quantum theories in terms of having a wavefunction ontology.17
They propose that in all quantum theories material entities are not represented by
the wavefunction but by some other mathematical entity in three-dimensional space
(or four-dimensional space-time) which they dubbed the primitive ontology (PO) of
the theory. The idea is that the ontology of a fundamental physical theory should
always be represented by a mathematical object in three-dimensional space, and
quantum mechanics is not an exception. Given that the wavefunction is a field in
configuration space, it cannot represent the ontology of quantum theory.
This view constitutes a realist approach, just like wavefunction realism. In
contrast with it, however, the configuration space problem never arises because
physical objects are not described by the wavefunction. In this way, not only the
pilot-wave theory but also the spontaneous collapse and many-worlds are ‘hidden
variable’ theories, in the sense that matter needs to be described by something else
(in three-dimensional space) and not by the wavefunction. For instance, primitive
ontologists talk about GRWm and GRWf as spontaneous localization theories with
a matter density m and an event (‘flash’) ontology f, as opposed to what they call
GRW0, namely the spontaneous localization theory with merely the wavefunction
(Allori et al. 2008).18 The proponents of this view therefore, contrarily to Bub
and Pitowsky, do not deny the first dogma: measurement processes are ‘regular’
physical processes analyzable in terms of their microscopic constituents (the PO).

15 With this I do not mean that there are no proposals but simply that these proposals are work-in-
progress, rather than full-blown solutions.
16 One may think that the fact that there is more than one solution to the measurement problem

implies that it has not really been solved. That is, one may think that the measurement problem
would be solved only if we had a unique answer to it. While this is certainly true, the sense
in which it has been solved (albeit not uncontroversially so) is that all its possible solutions are
fully developed, mature accounts, in contrast with the proposals to solve the configuration space
problem, which are only in their infancy. This is therefore the sense in which the configuration
space problem is more serious than the measurement problem.
17 Dürr et al. (1992), Allori et al. (2008), Allori (2013a, b).
18 Notice however, that they are hidden variable theories only in the sense that one cannot read

their ontology in the formalism of quantum mechanics, which is, in this approach, fundamentally
incomplete.
30 V. Allori

Nevertheless, like the IT interpretation, they reject the second dogma: they deny the
wavefunction representing physical objects. As we will see in Sect. 2.8, however,
they do not take the wavefunction as epistemic: this view rejects the second dogma
in the sense that the wavefunction is not material but not in the sense that it does not
represent some objective feature of the world.
In this framework, one of the motivations for the rejection of dogma 2 is the
conviction that the configuration space problem is a more serious problem than
the measurement problem. The measurement problem is the problem of making
sense of unobserved macroscopic superpositions which are consequences of the
formalism of the theory. I wish to argue now that one important reason why one
should drop the second dogma comes from the recognition that, even if the realist
solves the measurement problem, the corresponding realist theories would still face
the configuration space problem, which is arguably much harder to solve and which
can be entirely avoided if one rejects the second dogma.
First, take the wavefunction-only theories, like the spontaneous localization and
many-worlds theories. If one could keep the wavefunction in three-dimensional
space, then both these theories could identify a ‘particle’ as a localized three-
dimensional field given by the wavefunction (the superposition wavefunction
spontaneously localizes in the spontaneous localization theory, and decoherence
separates the different braches in Everett’s theory). This would solve the measure-
ment problem. However, these theories would need a wavefunction in configuration
space to be empirically adequate, and therefore one ends up facing the configuration
space problem. So, wavefunction realists solve one problem, but they have a new
one. In the case of the pilot-wave theory, moreover, the argument for the rejection
of the second dogma is even clearer, because the theory does not even solve the
measurement problem from the point of view of the wavefunction realist. In fact,
in this view matter would be made by both particles and waves. That is, the cat is
both made of waves and of particles, and because of its wave-like nature, she can
be in unobserved macroscopic superposition. The measurement problem is solved
only when one considers the theory a theory of particles only, without the wave-
counterpart representing matter. This is how indeed the theory is often discussed:
there are particles (which make up the cat), which are pushed around by the wave.
This is however still misleading because it assumes that some wave physically
exists, and this becomes problematical as soon as one realizes that this wave is
in configuration space: What kind of physical entity is it? How does it interact with
the particles? A better way of formulating the theory avoiding all this is to say that
cats are made of particles and the wavefunction is not material, therefore rejecting
the second dogma.19
Regardless, what’s so bad about the configuration space problem? According
to the proponents of the PO approach, it is extremely hard to solve, if not
unsolvable.20 Let us briefly see some reasons. One difficulty was recognized already

19 This is implicit in Dürr et al. (1992).

20 Allori et al. (2008), Allori (2013a, b).
2 Why Scientific Realists Should Reject the Second Dogma of Quantum Mechanics 31

by de Broglie, who noticed that, since for the wavefunction realist there are
no particles at the fundamental level, “it seems a little paradoxical to construct
a configuration space with the coordinates of points which do not exist” (de
Broglie 1927). Interestingly, this argument can be tracked down historically also
to Heisenberg, who very vividly said to Bloch (1976), referring to configuration
space realism: “Nonsense, [ . . . ] space is blue and birds fly through it”, to express
the ultimate unacceptability of building a theory in which there was no fundamental
three-dimensional space.21 Also, as already noticed, in wavefunction realism the
traditional explanatory schema of the behavior of macroscopic entities in terms of
microscopic ones needs to be heavily revised: contrarily to the classical case, tables
and chairs are not macroscopic three-dimensional objects composed of microscopic
three-dimensional particles; rather, as already noticed, they are macroscopic three-
dimensional objects ‘emerging from’ high-dimensional wavefunction. And this
problem has to be solved in absence of a truly compelling argument to do so. In
fact, what is the positive argument for dogma 2? The one we already saw is that
we should do what we did in Newtonian mechanics, namely extract the ontology
from the evolution equation of the theory. However here the situation is different.
Newton started off with an a priori metaphysical hypothesis and built his theory
around it, so there is no surprise that one has no trouble figuring out the ontology of
the theory. In contrasts, quantum mechanics was not developed like that. Rather, as it
is known, different formulations have been proposed to account for the experimental
data (Heisenberg matrix mechanics and Schrödinger wave mechanics) without
allowing ideas about ontology to take precedence. Just like Bub and Pitowsky claim,
before the solution of the measurement problem, the theory is a kinematic theory:
just like thermodynamics, it put constraints on possible experimental outcomes
without investigating how and why these outcomes come about. The main difference
between the PO and the IT approaches is that the former provides a framework of
constructive quantum theories based on the microscopic PO, which can be seen as
the dynamical counterpart of the latter, in which measurements are fundamental.22
Recently, a more compelling argument for wavefunction realism has been
defended, namely that wavefunction realism is the only picture in which the world
is local and separable.23 The idea is that, since fundamentally all that exists is the

21 Similar concerns have been expressed by Lorentz, who in a 1926 letter to Schrödinger wrote:
“If I had to choose now between your wave mechanics and the matrix mechanics, I would give
the preference to the former, because of its greater intuitive clarity, so long as one only has to
deal with the three coordinates x, y, z. If, however, there are more degrees of freedom, then I
cannot interpret the waves and vibrations physically, and I must therefore decide in favor of matrix
mechanics” (Przibram 1967). Similarly, Schrödinger wrote: “The direct interpretation of this wave
function of six variables in three-dimensional space meets, at any rate initially, with difficulties
of an abstract nature” (1926, p. 39). Again: “I am long past the stage where I thought that one
can consider the w-function as somehow a direct description of reality” (1935a). This is also a
concern heartfelt by Einstein, who expressed this view in many letters, e.g.: “The field in a many-
dimensional coordinate space does not smell like something real” (Einstein 1926).
22 For more on the comparison between the PO approach and the IT framework, see Dunlap (2015).
23 See Loewer (1996) and Ney (forthcoming) and references therein.
32 V. Allori

high dimensional wavefunction, there are no other facts that fail to be determined by
local facts about the wavefunction. First, let me notice that this is true also for the IT
account, in which what is real are the (three-dimensional) experimental outcomes.
However, while a more thorough analysis of this should be done somewhere else,
let me point out that it is an open question whether the notions of locality and
separability used in this context are the ones which have been of interest in the
literature. Einstein first proposed the EPR experiment to discard quantum mechanics
because it would show three-dimensional nonlocality. Arguably,24 Bell instead
proved that the locality assumption in EPR was false and that nonlocality is a fact of
nature. In other words, to show that the world is separable and local in configuration
space seems to give little comfort to people like Einstein who wanted a world which
is local and separable in three dimensional space.
Be that as it may, a further reason to reject the second dogma is that, if we
accept it and somehow we manage to solve the configuration space problem, still
we would lose important symmetries.25 However, this is not the case if we drop
the second dogma. In fact, the law of evolution of the wavefunction should no
longer be regarded as playing a central role in determining the symmetries of the
theory. Indeed, they are determined by the ontology, whatever it is (the PO in the
PO approaches, and measurement outcomes in the IT framework). If one assumes
that the wavefunction does not represent matter, and wants to keep the symmetry,
then one can allow the wavefunction to transform as to allow for this.26
Another reason to reject the second dogma, even assuming the configuration
space problem is solved, is the so-called problem of empirical incoherence. A
theory is said to be empirically incoherent when its truth undermines our empirical
justification for believing it to be true. Arguably, any theory gets confirmation by
spatiotemporal observations. However, wavefunction realism rejects space-time as
fundamental so that its fundamental entities are not spatiotemporal. Because of this,
our observations are not spatiotemporal observations, and thus provide no evidence
for the theory in the first place.27
Moreover, it seems to me that insisting on holding on to dogma 2 at all costs
would ultimately undermines the original realist motivation. In fact, the realist, in

24 See Bell and Gao (2016) and references therein for a discussion of Bell’s theorem and its
implications.
25 See Allori et al. (2008) and Allori (2018, 2019). For instance, quantum mechanics with

wavefunction ontology turns out not Galieli invariant. In fact the wavefunction is a scalar field
in configuration space, and as such will remain the same under a Galilean transformation:
ψ(r) → ψ(r − vt). In contrast,
invariance would require a more complicated transformation:
i
mvr− 1 mv 2 t
(r, t) → e  2
ψ (r − vt, t) . A similar argument can be drawn to argue that if one
accepts the second dogma then the theory is also not time reversal invariant.
26 See respectively Allori (2018, 2019) for arguments based on Galilean symmetry and time reversal

symmetry that the wavefunction is best seen as a projective ray in Hilbert space rather than a
physical field.
27 Barrett (1999), Healey (2002), Maudlin (2007). For responses, see Huggett and Wüthrich (2013)

and Ney (2015).

2 Why Scientific Realists Should Reject the Second Dogma of Quantum Mechanics 33

trying to make sense of the puzzles and paradoxes of quantum mechanics, as we

have seen, would try to find the ontology with methods that worked in the past. That
is, she will look at the evolution equation and she will identify the wavefunction as
the ontology of the theory. This move allows us to see the measurement problem.
To solve this problem the wavefunction realist develops theories with have an
evolution equation for an object in configuration space. Thus, she needs also to
solve the configuration space problem. Assuming a solution is possible, it will be
radically different from the familiar constructive theories since it would involve
emergence of the three-dimensional world from the multidimensional configuration
space, rather that the dynamical explanation of the three-dimensional macroscopic
data in terms of the three-dimensional microscopic ontology. However, the quest
for a dynamical explanation, typical of proponents of constructive theories, in
which macroscopic objects are thought of as composed of microscopic entities,
motivated the whole realist approach to quantum theory to start with. To put it in
another way, the wavefunction realist assumes that realist quantum theories (which
solve the measurement problem) are satisfactory as they are, and looks to their
fundamental equation to find the ontology. However the cost of assuming this
amounts to ending up with a theory which will certainly not involve a dynamical
explanation of the phenomena, which however is arguably the idea that motivated
the realist to solve the measurement problem to start with. In contrast the primitive
ontologist recognizes the disanalogy with classical mechanics (as mentioned earlier
in this section), in which the strategy of looking at the evolution equation to find
the ontology worked because Newton had already an ontology in mind when
he constructed the theory (in contrast with quantum theory). Thus, she rejects
dogma 2 by dropping the idea that the various realist theories are satisfactory as
they are28 : the wavefunction never describes matter, and therefore never provides
the complete description of a physical system, contra Bell’s characterization of
the measurement problem. Physical objects have a different ontology, which (to
avoid the configuration space problem) is postulated to be in three-dimensional
space. In this way, one can have a dynamical explanation of the phenomena in
terms of the fundamental ontology. In this sense, the PO approach provides with
constructive dynamical theories corresponding to the kinematic theory given by the
IT interpretation.
Before concluding this section, let me make some remarks. First, it is interesting
to note how the discussion above, in which we have motivated the rejection of
the second dogma with multiple arguments, provides a response to Timpson, who
complained that Bub and Pitowsky provide no reason why one would be unhappy
with dogma 2.
In addition, let me mention another theory which accepts the second dogma, as
well as the first, which has been proposed to avoid the configurations space problem.
The idea is to think as the wavefunction as a multi-field in three-dimensional

28 Notice that there is nothing strange in this: the wavefunction realist did not accept quantum
mechanics as it is because it suffered from the measurement problem, and in order to solve it, she
finds it acceptable to, for instance, change the theory by modifying the Schrödinger equation.
34 V. Allori

field. That is, the wavefunction is a ‘poly-wave’ on three-dimensional space, a

generalization of an ordinary classical field: as a classical field specifies a definite
field value for each location in three-dimensional space, the multi-field, given an
N-particle system, specifies a value for the N-tuple of points in three-dimensional
space.29 Even if it has the advantage of having no configuration space problem
without dropping the second dogma, however this approach is problematical
because it loses important symmetry properties, given that the multi-field transforms
differently from how a classical field would (Belot 2012).30
On a different note, observe that the reasons I provided for the rejection of
dogma 2 are not necessarily connected with dogma 1. On one hand, one can
reject it with dogma 2, as in the IT approach, and have a macroscopic ontology
of measurement results without having the measurement problem. The existence of
the IT interpretation, which I think is best seen as a refinement of Bohr’s original
account, is thus the counterexample to the claim that a scientific realist cannot solve
the measurement problem by rejecting both dogmas. On the other hand, PO theories
solve the measurement problem by rejecting only the second dogma. Criticisms to
theories which reject the first dogma has been put forward. Most notably Bell (1987)
argues that by taking measurements as primitive one introduces a fundamental
vagueness in the theory connected with characterizing what a measurement is. This,
it is argued, makes the theory hardly satisfactory. Moreover, Egg (2018) argues that
rejecting the first dogma is incompatible with standard accounts of scientific realism,
according to which reality is more than what we can observe or measure. In the
next section we will discuss in a little more detail the different options open to the
realist, and the factors which are likely to be relevant in making a choice among the
alternatives.

2.7 The Dogmas, Scientific Realism, and the Role

of Explanation

Let’s summarize some of the points made so far, and draw some partial conclusions.
A scientific realist thinks it is possible to explain the manifest image of our senses
in terms of the scientific image given to us by our best theories (Sellars 1962). We
have different possibilities:
1. Reject dogma 2, thus assume a microscopic ontology in three-dimensions, as
advised by the PO approach;

29 Forrest (1988), Belot (2012), Hubert and Romano (2018).

30 Moreover, a new formulation of the pilot-wave theory has been proposed (Norsen 2010) in
which to each particle is associated a three-dimensional filed given by a conditional wave-function,
namely a function ψ defined by the wave-function of the universe
once the positions of all the
other particles in the universe Y(t) are fixed: ψ t (x):=
(x, Y(t)). However, to recover the correct
trajectories one needs to add infinitely many fields, and this makes the theory hardly satisfactory.
2 Why Scientific Realists Should Reject the Second Dogma of Quantum Mechanics 35

2. Reject both dogmas, thus assume a macroscopic ontology in three-dimensions,

as advised by the IT interpretation; and
3. Accept both dogmas, thus assume a non-three-dimensional ontology, as in
wavefunction realism.31
In the first case, one can provide a constructive and dynamical explanation of the
manifest image, both of the experimental results and the properties of macroscopic
objects. In the second case, instead, the only thing to explain is why certain
statistical patterns are observed, and this is done using kinematical constrain on
the possible physical events, not in a dynamical way. In the third case the situation
is more complicated: one needs to explain both the empirical statistics and the
‘emergence’ of macroscopic objects in a way which is neither (obviously) kinematic
nor dynamics, neither constructive nor using principles. No such account has been
proven so far as being uncontroversial or fully developed. Because of this, and the
other problems discussed in the previous section, all other things being equal, at this
stage one can argue that the first two alternatives are to be preferred over the third.
So, which of the first two should we chose? This is going to depend on whether
one prefers dynamical to kinematic explanations. That is, Bub and Pitowsky find
option 1 unnecessary, while the proponents of the PO approach find 2 lacking
explanatory power. If so, we have found a new role for the kinematic/dynamic
distinction. In the original argument by Bub and Pitowsky the distinction came in the
first step of the argument to reject the first dogma on the basis of which dogma 2 was
later rejected. Because the superiority of kinetic theories is a controversial matter,
however, I think this argument is not likely to convince anyone already inclined to
prefer dynamical theories to drop dogma 2. In contrast, in the argument I proposed,
the reasoning against dogma 2 is independent of this distinction, which only plays a
role in selecting which of the alternatives theories which already reject dogma 2.

2.8 On the Status of the Wavefunction

Before discussing the objections to the IT interpretation based on the epistemic

account of the wavefunction, let us take another step back. The different accounts of
the nature of the wavefunction are usually classified as either ‘ontic’ or ‘epistemic’.
Let’s see what this means, and review some of the main objections to the ontic

31 Interestingly,
there’s logically a fourth option, namely a theory which rejects dogma 1 without
rejecting dogma 2. This would be a theory in measurement outcomes are left unanalyzed and the
wavefunction represents physical objects. I am not sure who would hold such a view, but it seems to
me that such a theory would presumably not appeal many people: while it solves the measurement
problem (by rejecting dogma 1) it does not solve the configuration space problem (since it accepts
dogma 2). Moreover, this theory leaves a mystery what the relation between the wavefunction and
measurement processes is. Regardless, it nicely shows the relation between the two dogmas, with
dogma 2 being in my opinion the one responsible for the problems for the quantum realist.
36 V. Allori

accounts, while we will leave the criticism to the epistemic accounts to the next
section.
An account is said to be ontic when the wavefunction (or the density matrix
which may also define the quantum state) represents some sort of objective feature
of the world. In contrast, an account is epistemic if the wavefunction (or the
density matrix) has to do with our knowledge of the physical system. Each
view has their own sub-classification. As we have seen, wavefunction realism
and the wavefunction as multi-field are ontic approaches in that they regard the
wavefunction as a physical field. However, there are other approaches that could be
labelled as ontic which do not have this feature. One of these ontic interpretations
is the view according to which the wavefunction is a law of nature. This is a
peculiar account because it has been formulated in the PO framework, which denies
the second dogma. Roughly put, the idea is that since the wavefunction does not
represent matter but has the role of governing the behavior of matter, then the
wavefunction has a nomological character and thus is best understood as a law of
nature.32 Notice that this account of the wavefunction is still ontic (the wavefunction
expresses some nomic feature of the world) even if the second dogma is rejected
(the wavefunction does not represent material entities), and thus one can dub this
view as ‘ψ-non-material’ rather than ‘ψ-ontic’, which does not seem sufficiently
specific (another non-material view is the one according to which the wavefunction
is a property, see below). Perhaps the most serious problem33 for this approach is
that one usually thinks of laws as time-independent entities, while the wavefunction
evolves itself in time.34 Another challenge for the view is that while laws are taken
to be unique, the wavefunction is not. Replies have been provided, relying on a
future quantum cosmology in which the wave function would be static.35 However,
they have been received with skepticism, given that one would want to get a handle
on the nature of the wavefunction in the current situation, without having to wait for
a future theory of quantum cosmology.
Perhaps a better way of capturing this idea can be found in a Humean framework,
regarding the wavefunction as part of the Humean mosaic.36 Nonetheless, many
still find the account wanting, most prominently because they find Humeanism with
respect to laws misguided for other reasons. A different but connected approach that
can be thought as nomological is to think of the wave function as a dispositional
property or more generally as a property of the ontology.37 Not surprisingly, even if
these approaches do not have the time-dependence problem, they nevertheless are
not immune to criticisms, given that the dispositional properties are notoriously a

32 Dürr et al. (1992, 1997), Goldstein and Zanghì (2013).

33 For more objections, see Belot (2012), Callender (2015), Esfeld et al. (2014), Suárez (2015).
34 Brown and Wallace (2005).
35 Goldstein and Teufel (2001).
36 Miller (2014), Esfeld (2014), Callender (2015), and Bhogal and Perry (2017).
37 Suàrez (2007, 2015), Monton (2013), Esfeld et al. (2014), and Gao (2014).
2 Why Scientific Realists Should Reject the Second Dogma of Quantum Mechanics 37

tough nut to crack and moreover the wavefunction does not behave like a regular
property.38
Moving on to the epistemic approaches, they have in common the idea that the
wavefunction should be understood as describing our incomplete knowledge of the
physical state, rather than the physical state itself. One of the motivations for the
epistemic approaches is that they immediately eliminate the measurement problem
altogether. In fact, schematically, in a Schrödinger cat type of experiment, the cat is
never in a superposition state; rather, it is the experimenter who does not know what
its state is. When I open the box and find the cat alive, I simply have updated my
knowledge about the status of the cat, who now I know to always have been alive
during the experiment.
The epistemic approaches can be distinguished into the ‘purely epistemic,’ or
neo-Copenhagen, accounts, and the ones that invoke hidden variables, which one
can dub ‘realist-epistemic’ approaches. The prototype of the latter type is Einstein’s
original ‘ignorance’ or ‘statistical’ interpretation of the wavefunction (Einstein
1949a), according to which the current theory is fundamentally incomplete: there are
some hidden variables which describe the reality under the phenomena, and whose
behavior is statistically well described by the wavefunction.39 For a review of these
epistemic approaches, see Ballentine (1970).40 In contrast, the neo-Copenhagen
approaches reject that the above mentioned hidden variables exist, or are needed.
The wavefunction thus describes our knowledge of measurement outcomes, rather
than of an underlying reality. This view has a long story that goes back to some
of the founding fathers of the Copenhagen school as one can see, for instance,
in Heisenberg (1958) and in Peierls (1991), who writes that the wavefunction
“represents our knowledge of the system we are trying to describe.” Similarly, Bub
and Pitowsky write: “quantum state is a credence function, a bookkeeping device
for keeping track of probabilities”.41
A crucial difference between the ontic and the epistemic approaches is that while
in the former the description provided by the wavefunction is objective, in the
latter it is not. This is due to the fact that different agents, or observers, may have
different information about the same physical system. Thus, many epistemic states
may correspond to the same ontic state.

38 For an assessment, see Suàrez (2015).

39 Einstein was convinced that “assuming the success of efforts to accomplish a complete physical
description, the statistical quantum theory would, within the framework of future physics, take an
approximately analogous position to the statistical mechanics within the framework of classical
mechanics. I am rather firmly convinced that the development of theoretical physics will be of this
type; but the path will be lengthy and difficult” (Einstein 1949a, b).
40 For a more modern approach, see Spekkens (2007).
41 Other neo-Copenhagen approaches include Bayesian approaches (Fuchs and Peres 2000; Fuchs

2002; Fuchs and Schack 2009, 2010, Caves et al. 2002a, 2002b, 2007), pragmatist approaches
(Healey 2012, 2015, 2017; Friederich 2015), and relational ones (Rovelli 1996).
38 V. Allori

2.9 Objections Based on the Wavefunction Being Epistemic

As we anticipated, since the IT interpretation rejects the first dogma and takes
measurements as primitive and fundamental, one is left with the question about
what the wavefunction is. According to Bub and Pitowsky, the wavefunction (or the
density matrix) is not fundamental. Thus, they spend not too much time in discussing
what the nature of the wavefunction is, and Timpson (2010) finds this troublesome.
However, it seems natural for them to consider the wavefunction as epistemic, given
their overall approach: the wavefunction adds no empirical content to the theory,
and thus is best seen not as a description of quantum systems but rather as reflecting
the assigning agents’ epistemic relations to the systems.
One epistemic approach which shows important similarity with Bub and
Pitowsky’s view is the one put forward by Healey (2012, 2015, 2017). According
to him, physical situations are not described by the wavefunction itself. Rather
the wavefunction prescribes how our epistemic attitudes towards claims about
physical situations should be. For instance, in a particles interference experiment, an
experimenter uses the Born rule to generate the probability that the particle is located
in some region. The wavefunction therefore prescribes what the experimenter
should believe about particle location. This view is epistemic in the sense that
the wavefunction provides a guide to the experimenter’s beliefs. Moreover, it is
pragmatic in the sense that concerns about explanation are taken to be prior to
representational ones. In this sense, the wavefunction has an explanatory role.
This explanatory role is also shared by Bub and Pitowsky’s account, since in their
account the wavefunction allows the definition of the principles of the theory which
constrain the possible physical phenomena, including interference.
Anyway, because of its epistemic approach, the IT interpretation has received
many criticisms, and the situation remains extremely controversial.42 Let me
review what the main criticisms are. First of all, it has been argued that the
epistemic approaches cannot explain the interference phenomena in the two-slit
experiment for particles: how can our ignorance about certain details can make some
phenomenon such as interference happen? It seems that some objective feature of
the world needs to exist in order to explain the interference fringes, and the obvious
candidate is the wavefunction (Leifer 2014; Norsen 2017). Moreover, in particle
experiments with the Mach-Zender interferometer outcomes change depending on
whether we know or not which path the particle has taken. This seems incompatible
with the epistemic view: if the wavefunction represent our ignorance about which
path has been taken, then coming to know that would change the wavefunction but
would not produce any physical effect (Bricmont 2016).
Moreover, the so-called no-go theorems for hidden variables (starting from the
ones of von Neumann 1932, and Bell 1987) have been claimed to show that the
realist-epistemic view is untenable, so that they can be considered no-go theorems

42 See Gao (2017) for a review of the objections, and Leifer (2014) for a defense.
2 Why Scientific Realists Should Reject the Second Dogma of Quantum Mechanics 39

for this type of approach. In fact, if the wavefunction represents our ignorance of
an underlying reality in which matter has some property A, say, spin in a given
direction, then the wavefunction provides the statistical distribution of v(A), the
values of A. Indeed, if this interpretation is true, v(A) exists for more than one
property, since it would be arbitrary to say that spin exists only on certain directions.
However, the no-hidden variable theorems show that mathematically no v(A) can
agree with the quantum predictions. That is, one cannot introduce hidden variables
to all observables at once.43
Recently, a new theorem called PBR theorem, from the initials of its proponents
(Pusey et al. 2012) has been proposed as a no-go theorem for the epistemic theories
in general. This theorem is supposed to show that, in line of the strategy used by
the other no-go theorems, epistemic approaches are in conflict with the statistical
predictions of quantum mechanics by showing that if the wavefunction is epistemic,
then it requires the existence of certain relations which are mathematically impossi-
ble. Therefore, if quantum mechanics is empirically adequate, then the wavefunction
must be ontic (either material, or not).
This objections have been subject to an extensive examination, and proponents of
the epistemic views have proposed several replies, which are however in the eyes of
many still tentative.44 Bub and Pitowsky respond that they can explain interference
as a phenomenon constrained by the structure of Hilbert space, more specifically by
the principle of ‘no universal broadcasting.’ Moreover, PBR’s conclusion, as well as
the conclusions of the other no-go theorems, applies only to theories that assume an
underlying reality, which they reject. However, these answers appear unsatisfactory
to someone who would want to know why these phenomena obtain, rather than
having a mere description. So, without entering too much in the merit of these
objections, in the view of many they pose a serious challenge to the epistemic views,
and, accordingly, to the IT interpretation. Therefore, one may wonder whether it is
possible to deny the second dogma while staying away from the epistemic views
without endorsing the non-material interpretations discussed in Sect. 2.8, given that
they suffer from serious objections too. This is the enterprise I will take on in the
next section.

2.10 The Wavefunction Is as the Wavefunction Does

We have reviewed in the previous sections the various non-material accounts of the
wavefunction (the nomological approach, which is non-material but ontic, and the
epistemic ones) and their criticisms. In this section, I propose a new account of the
wavefunction based on functionalism, which I argue is better than the alternatives
and could be adopted both by the proponents of the PO approach and by Bub and
Pitowsky.

43 Notice that the pilot-wave theory circumvents this theorem because some, but not all, experi-
ments reveal pre-existing properties (for more on this, see Bricmont 2016; Norsen 2017).
44 Again, see Leifer (2014), Gao (2017), and references therein.
40 V. Allori

Functionalism, broadly speaking, is the view that certain entities can be ‘reduced’
to their functional role. To use a powerful slogan, ‘a table is as a table does.’
Strategies with a functionalist core have been used in the philosophy of mind for
a long time (starting from Putnam 1960), and they have been recently used also in
philosophy of physics.
For instance Knox (2013, 2014) argues that spacetime can be functionalized
in the classical theory of gravity. That is, spacetime can be thought of as non-
fundamental (emergent) in virtue of the fact that spacetime plays the role of
defining inertial frames. Interestingly, Albert (2015) defends wavefunction realism
using his own brand of functionalism. He in fact argues that ordinary three-
dimensional objects are first functionalized in terms of their causal roles, and
the wavefunction dynamically uses these relations in a way which gives rise to
the empirical evidence.45 Lam and Wüthrich (2017) use functionalist strategies in
quantum gravity. There are many proposals for how to articulate a quantum theory
of gravity, however it has been argued that in all these theories spacetime is not
fundamental, but rather it is emergent from some non-spatiotemporal structures
(Huggett and Wuthrich forthcoming). They argue that in order for spacetime to
emerge it suffices to recover only those features which are functionally relevant
in producing observations. However, not surprisingly, this approach faces the same
objections of wavefunction realism. An objection to this approach is that it is unclear
how space and time can indeed emerge from a fundamental non-spatiotemporal
ontology: first, it is unclear what a non-spatiotemporal fundamental could be;
second, it is doubtful whether the right notion of emergence is the one proposed
(Lam and Esfeld 2013). In addition, there is the already mentioned problem of
empirical incoherence, namely that the theory undermines its own confirmation.
Arguably, any theory gets confirmation by spatiotemporal observations. However, a
theory in which spacetime is functionalized is a theory whose fundamental entities
are not spatiotemporal. Because of this, our observations are not spatiotemporal
observations, and thus provide no evidence for the theory in the first place.46
As anticipated, my idea is that one should functionally define the wavefunction,
similarly as how people wish to functionalize space, space-time and material
objects, and define it in terms of the role it plays in the theory. That is, the
wavefunction is as the wavefunction does. I am going to show that this view captures
the appeals of the other accounts without falling prey to their objections. In other
words, by thinking of the wavefunction functionally one can capture the intuitions
that the wavefunction is law-like (as proposed by the nomological approach), and
that it is explanatory rather than representational (as proposed by the epistemic and
pragmatic approaches) avoiding the problem of time-dependence, non-uniqueness,
and no-go theorems.
The wavefunction plays different but interconnected roles in realist quantum
theories which reject the second dogma, which however all contribute to the same

45 Ney (2017) provides a criticism and discusses her own alternative. See also Ney (forthcoming).
46 For responses, see Huggett and Wüthrich (2013), and Ney (2015, forthcoming).
2 Why Scientific Realists Should Reject the Second Dogma of Quantum Mechanics 41

goal: to reproduce the empirical data. First, the wavefunction plays a nomological
role. This is straightforward in the case of PO theories, where it is one of the
ingredients necessary to write the dynamics for the PO, which in turn constructively
accounts for the measurement outcomes.47 In the case of the IT interpretation, the
dynamics plays no role, but the wavefunction can be seen as nomological in the
sense that allows the principles or the kinematic constraints of the theory to be
expressed, which in turn limit the physically possible phenomena and thus explain
the data.
Connected to this, the wavefunction does not represent matter but it has an
explanatory role: it is a necessary ingredient to account for the experimental results.
As we have seen, the PO approach and the IT interpretation differ in what counts as
explanatory, the former favoring a dynamical account, and the latter a kinematic one.
Accordingly, the explanatory role of the wavefunction in these theories is different.
In the PO approach the wavefunction helps defining the dynamics for the PO, and
in this way accounts for the empirical results constructively as macroscopically
produced by the ‘trajectories’ of the microscopic PO. In the IT interpretation the
wavefunction provides the statistics of the measurement outcomes via the Born rule,
therefore explaining the probabilistic data by imposing kinematical constraints.
Notice that this explanatory role allows to be pragmatic about the wavefunction:
it is a useful object to adequately recover the empirical predictions. Pragmatically,
both in the PO approach and in the IT framework one uses the wavefunction to
generate the probability outcomes given by the Born rule, but one could have chosen
another object, like a density matrix, or the same object evolving according to
another evolution equation.48 The fact that quantum mechanics is formulated in
terms of the wavefunction is contingent to other super-empirical considerations
like for example simplicity and explanatory power: it is the simplest and most
explanatory choice one could have made.
This pragmatic take is compatible also with the nomological role of the wave-
function. In fact, as we have seen, the wavefunction is not a law of nature,
strictly speaking, rather it is merely an ingredient in the law. Because of this the
wavefunction does not have to be unique, like potentials are not unique, and can be
dispensable: other entities, such as the density matrix, can play its role. In this way,
the wavefunction plays a nomological role, but avoiding the usual objections to the
nomological view.
Moreover, notice that this approach, even if pragmatic, is not epistemic: the
wavefunction does not have to do with our knowledge of the system (contra Bub and
Pitowsky’s original proposal). As a consequence, it does not suffer from the usual
problems of the epistemic approaches. Moreover, in the case of the IT approach, this

47 In the context of the pilot-wave theory, the wavefunction can be taken as a force or a potential,
but one should not think of it as material (even if Belousek 2003 argues otherwise).
48 In the PO framework, for instance, see Allori et al. (2008) for unfamiliar formulations of the

pilot-wave theory and the spontaneous collapse theory respectively with a ‘collapsing’ wave
function and a non-collapsing one respectively.
42 V. Allori

view remains true to the ideas which motivated the interpretation, namely to reject
both dogmas, and fits well with their ideas of quantum probability.
Also, I wish to remark that, since all the theories we are considering reject the
second dogma, this view also avoids the objection of empirical incoherence of other
functionalist approaches. In fact the problem originates only if the fundamental
entities of the theory are not spatiotemporal, and this is not the case in the IT
frameworks (in which measurements outcomes are fundamental and are in three-
dimensional space) and in the PO approach (in which the PO is a spatiotemporal
entity).
To summarize, the wavefunction plays distinctive but interconnected roles which
act together and allow to recover the empirical data. First, the wavefunction is
an ingredient in the laws of nature, and as such contributes in determining the
trajectories of matter. As such, it has an explanatory role in accounting of the
experimental results, even if it is not representational one. Because of this, the
wavefunction can be thought in pragmatic terms without considering it connected to
our knowledge of the system. Formulated in this way, the view has the advantages
of capturing the main motivations for the other views without falling prey of their
main objections:
1. The wavefunction has a nomological role, without being a law; so it does not
matter whether it is time-evolving or unique;
2. The wavefunction has an explanatory role without necessarily being uniquely
defined or connected to the notion of dynamical explanation;
3. The wavefunction has a pragmatic role in recovering the empirical data without
being epistemic, thereby avoiding the interference problem and the no-go
theorems linked with the epistemic view.
These roles combine together to define the wavefunction functionally: the
wavefunction is whatever function it plays, namely to recover the experimental
results. This role can be accomplished either directly, in the IT framework, or
indirectly, through the ‘trajectories’ of the PO, but always remaining in spacetime,
therefore bypassing the configuration space problem and the worry of empirical
incoherence. This account of the wavefunction thus is arguably better than the other
taken individually, and accordingly is the one that the proponents of realist theories
which reject the second dogma should adopt. In particular, it can be adopted by the
proponents of the PO approach, as well as by Bub and Pitowsky to strengthen their
views.

2.11 Conclusion

As we have seen, the IT interpretation rejects the two dogmas of quantum

mechanics. First, it rejects dogma 1. Using the CBH theorem, which kinematically
constrains measurements outcomes, they see quantum theory as a theory with
measurements as primitive. Then, what is the wavefunction? The IT interpretation
2 Why Scientific Realists Should Reject the Second Dogma of Quantum Mechanics 43

naturally interpret it as epistemic, therefore rejecting the second dogma. It is argued

that both dogmas give rise to the measurement problem, and that by rejecting them
both it becomes a pseudo-problem.
Because of this sequence of arguments the IT interpretation has the following
two features that could be seen as shortcomings:
1. The rejection of dogma1 is based on the superiority of kinematic theories, which
is controversial;
2. The rejection of dogma 2 makes them endorse the epistemic view, which is also
considered problematical.
My aim in this paper was:
A. To show that any realist should reject dogma 2, even if she may or may not reject
dogma 1, and
B. To provide a functionalist approach to the wavefunction which captures some
of the motivations of the other accounts, including IT, while avoiding their main
objections.
Aside from standing for themselves, in doing A, I also have provided the IT
interpretation with tools to avoid the first set of objections; while in doing B, a way
to reject the second. In fact, first I have argued that Bub and Pitowsky should have
rejected dogma 2 independently of dogma 1 (on the basis of the configuration space
problem, the loss of symmetries, and empirical incoherence), thereby completely
bypassing the connection with the alleged superiority of kinematic theories. Indeed,
my reasons for rejecting dogma 2 are less contentious in the sense that even
those who endorse dogma 2 (such as the wavefunction realists) recognize them
as problems for the tenability of dogma 2. After rejecting dogma 2 for these
reasons, someone may also drop dogma 1 to make quantum mechanics a kinematic
framework (if they prefer them for independent reasons). However, one may
decide otherwise: a realist who prefers constructive theories based on a dynamical
explanation could still keep dogma 1. Thus, if the IT proponents follow the route
proposed here (drop dogma 2 first) then they can avoid their first type of challenges.
As we have seen, the second set of objections comes from them endorsing the
epistemic view. However, my functionalist approach to the wavefunction, which
has merits in its own right, is compatible with dropping both dogmas without
committing us to an epistemic view of the wavefunction. Because of this, it can
be adopted by Bub and Pitowsky to avoid the second set of problems.

Acknowledgements Thank you to Meir Hemmo and Orly Shenker for inviting me to contribute to
this volume. Also, I am very grateful to Laura Felline for her constructive comments on an earlier
version of this paper, and to the participants of the International Workshop on “The Meaning of the
Wave Function”, Shanxi University, Taiyuan, China, held in October 2018, for important feedback
on the second part of this paper.
44 V. Allori

References

Albert, D. Z. (1996). Elementary quantum metaphysics. In J. T. Cushing, A. Fine, & S. Goldstein

(Eds.), Bohmian mechanics and quantum theory: An appraisal (pp. 277–284). Dordrecht:
Kluwer.
Albert, D. Z. (2013). Wave function realism. In D. Z. Albert & A. Ney (Eds.), The wave function:
Essay on the metaphysics of quantum mechanics (pp. 52–57). New York: Oxford University
Press.
Albert, D. Z. (2015). After physics. Cambridge: Harvard University Press.
Allori, V. (2013a). On the metaphysics of quantum mechanics. In S. Lebihan (Ed.), Precis de la
Philosophie de la Physique. Paris: Vuibert: 116–151.
Allori, V. (2013b). Primitive ontology and the structure of fundamental physical theories. In D. Z.
Albert & A. Ney (Eds.), The wave-function: Essays in the metaphysics of quantum mechanics
(pp. 58–75). New York: Oxford University Press.
Allori, V. (2018). A new argument for the Nomological interpretation of the wave function: The
Galilean group and the classical limit of nonrelativistic quantum mechanics. International
Studies in the Philosophy of Science, 31(2), 177–188.
Allori, V. (2019). Quantum mechanics, time and ontology. Studies in History and Philosophy of
Modern Physics, 66, 145–154.
Allori, V., Goldstein, S., Tumulka, R., & Zanghì, N. (2008). On the common structure of Bohmian
mechanics and the Ghirardi-Rimini-Weber Theory. The British Journal for the Philosophy of
Science, 59(3), 353–389.
Bacciagaluppi, G., & Valentini, A. (2009). Quantum theory at the crossroads: Reconsidering the
1927 Solvay conference. Cambridge: Cambridge University Press.
Balashov, Y., & Janssen, M. (2003). Critical notice: Presentism and relativity. The British Journal
for the Philosophy of Science, 54, 327–346.
Ballentine, L. E. (1970). The statistical interpretation of quantum mechanics. Reviews of Modern
Physics, 42(4), 358–381.
Barrett, J. A. (1999). The quantum mechanics of minds and worlds. New York: Oxford University
Press.
Bell, J. S. (1976). How to teach special relativity. Progress in Scientific Culture, 1. Reprinted in
Bell (1987), pp. 67–80.
Bell, J. S. (1987). Speakable and unspeakable in quantum mechanics. Cambridge: Cambridge
University Press.
Bell, M., & Gao, S. (2016). Quantum nonlocality and reality, 50 years of Bell’s theorem.
Cambridge: Cambridge University Press.
Belot, G. (2012). Quantum states for primitive ontologists. European Journal for Philosophy of
Science, 2(1), 67–83.
Belousek, D. W. (2003). Formalism, ontology and methodology in Bohmian mechanics. Founda-
tions of Science, 8(2), 109–172.
Bhogal, H., & Perry, Z. (2017). What the Humean should say about entanglement. Noûs, 1, 74–94.
Bloch, F. (1976). Heisenberg and the early days of quantum mechanics. Physics Today.
Bohm, D. (1952). A suggested interpretation of the quantum theory in terms of ‘hidden variables’
I & II. Physical Review, 85(2), 166–179. 180–193.
Bricmont, J. (2016). Making sense of quantum mechanics. Cham: Springer.
Brown, H. R. (2005). Physical relativity: Space-time structure from a dynamical perspective
(forthcoming). Oxford: Oxford University Press.
Brown, H. R., & Pooley, O. (2004). Minkowski space-time: A glorious non-entity. In D. Dieks
(Ed.), The ontology of spacetime (pp. 67–89). Amsterdam: Elsevier.
Brown, H. R., & Timpson, C. (2006). Why special relativity should not be a template for a
fundamental reformulation of quantum mechanics. In W. Demopoulos & I. Pitowsky (Eds.),
Physical theory and its interpretation: Essays in honor of Jeffrey Bub (pp. 29–42). Dordrecht:
Springer.
2 Why Scientific Realists Should Reject the Second Dogma of Quantum Mechanics 45

Brown, H. R., & Wallace, D. (2005). Solving the measurement problem: de Broglie-Bohm loses
out to Everett. Foundations of Physics, 35, 517–540.
Bub, J. (2004). Why the quantum? Studies in the History and Philosophy of Modern Physics, 35,
241–266.
Bub, J. (2005). Quantum mechanics is about quantum information. Foundations of Physics, 35(4),
541–560.
Bub, J. (2007). Quantum probabilities as degrees of belief. Studies in History and Philosophy of
Modern Physics, 39, 232–254.
Bub, J., & Pitowsky, I. (2010). Two dogmas about quantum mechanics. In S. Saunders, J. Barrett,
A. Kent, & D. Wallace (Eds.), Many worlds?: Everett, quantum theory, and reality (pp. 433–
459). Oxford: Oxford University Press.
Callender, C. (2015). One world, one Beable. Synthese, 192, 3153–3177.
Caves, C. M., Fuchs, C. A., & Schack, R. (2002a). Quantum probabilities as Bayesian probabilities.
Physical Review A, 65, 022305.
Caves, C. M., Fuchs, C. A., & Schack, R. (2002b). Unknown quantum states: The quantum de
Finetti representation. Journal of Mathematical Physics, 44, 4537–4559.
Caves, C. M., Fuchs, C. A., & Schack, R. (2007). Subjective probability and quantum certainty.
Studies in History and Philosophy of Modern Physics, 38(2), 255–274.
Clifton, R., Bub, J., & Halvorson, H. (2003). Characterizing quantum theory in terms of
information theoretic constraints. Foundations of Physics, 33(11), 1561.
de Broglie, L. (1927). La Nouvelle Dynamique des Quanta. In: Solvay conference, Electrons
et Photons, translated in G. Bacciagaluppi and A. Valentini (2009) Quantum theory at the
crossroads: Reconsidering the 1927 Solvay conference (pp. 341–371). Cambridge: Cambridge
University Press.
Dunlap, L. (2015). On the common structure of the primitive ontology approach and information-
theoretic interpretation of quantum theory. Topoi, 34(2), 359–367.
Dürr, D., Goldstein, S., & Zanghí, N. (1992). Quantum equilibrium and the origin of absolute
uncertainty. Journal of Statistical Physics, 67, 843–907.
Dürr, D., Goldstein, S., & Zanghì, N. (1997). Bohmian mechanics and the meaning of the wave
function. In R. S. Cohen, M. Horne, & J. Stachel (Eds.), Experimental metaphysics — Quantum
mechanical studies for Abner Shimony (Vol. 1: Boston Studies in the Philosophy of Science)
(Vol. 193, pp. 25–38). Boston: Kluwer Academic Publishers.
Egg, M. (2018). Dissolving the measurement problem is not an option for realists. Studies in
History and Philosophy of Science Part B: Studies in History and Philosophy of Modern
Physics, 66, 62–68.
Einstein, A. (1919). What is the theory of relativity? The London Times. Reprinted in Einstein, A.
(1982). Ideas and opinions (pp. 227–232). New York: Crown Publishers, Inc.
Einstein, A. (1926). Einstein to Paul Ehrenfest, 18 June, 1926, EA 10–138. Translated in Howard
(1990), p. 83.
Einstein, A. (1949a). Autobiographical notes. In P. A. Schilpp (Ed.), Albert Einstein: Philosopher-
scientist. Evanston, IL: The Library of Living Philosophers.
Einstein, A. (1949b). Reply to criticisms. In P. A. Schilpp (Ed.), Albert Einstein: Philosopher-
scientist. Evanston, IL, The Library of Living Philosophers.
Esfeld, M. A. (2014). Quantum Humeanism, or: Physicalism without properties. The Philosophical
Quarterly, 64, 453–470.
Esfeld, M. A., Lazarovici, D., Hubert, M., & Dürr, D. (2014). The ontology of Bohmian mechanics.
The British Journal for the Philosophy of Science, 65, 773–796.
Everett, H., III. (1957). ‘Relative state’ formulation of quantum mechanics. Reviews of Modern
Physics, 29, 454–462.
Felline, L. (2011). Scientific explanation between principle and constructive theories. Philosophy
of Science, 78(5), 989–1000.
Flores, F. (1999). Einstein’s theory of theories and types of theoretical explanation. International
Studies in the Philosophy of Science, 13(2), 123–134.
Forrest, P. (1988). Quantum metaphysics. Oxford: Blackwell.
46 V. Allori

Friederich, S. (2015). Interpreting quantum theory: A therapeutic approach. Basingstoke: Palgrave

Macmillan.
Friedman, M. (1974). Explanation and scientific understanding. Journal of Philosophy, 71, 5–19.
Fuchs, C. A. (2002). Quantum mechanics as quantum information (and only a little more). In A.
Khrennikov (Ed.), Quantum theory: Reconsideration of foundations. Växjö: Vaxjo University
Press.
Fuchs, C. A., & Peres, A. (2000). Quantum theory needs no ‘interpretation’. Physics Today, 53(3),
70–71.
Fuchs, C. A., & Schack, R. (2009). Quantum-Bayesian coherence. Reviews of Modern Physics, 85,
1693.
Fuchs, C. A., & Schack, R. (2010). A quantum-Bayesian route to quantum-state space. Foundations
of Physics, 41(3).
Gao, S. (2014). Reality and meaning of the wave function. In S. Gao (Ed.), Protective measurement
and quantum reality: Toward a new understanding of quantum mechanics (pp. 211–229).
Cambridge: Cambridge University Press.
Gao, S. (2017). The meaning of the wave function: In search of the ontology of quantum mechanics.
Cambridge: Cambridge University Press.
Ghirardi, G. C., Rimini, A., & Weber, T. (1986). Unified dynamics for microscopic and macro-
scopic systems. Physical Review D, 34, 470.
Goldstein, S., & Teufel, S. (2001). Quantum Spacetime without observers: Ontological clarity and
the conceptual foundations of quantum gravity. In C. Callender & N. Huggett (Eds.), Physics
meets philosophy at the Planck scale. Cambridge: Cambridge University Press.
Goldstein, S., & Zanghì, N. (2013). Reality and the role of the wave function in quantum theory.
In D. Albert & A. Ney (Eds.), The wave-function: Essays in the metaphysics of quantum
mechanics (pp. 91–109). New York: Oxford University Press.
Healey, R. (2002). Can physics coherently deny the reality of time? Royal Institute of Philosophy
Supplement, 50, 293–316.
Healey, R. (2012). Quantum theory: A pragmatist approach. The British Journal for the Philosophy
of Science, 63(4), 729–771.
Healey, R. (2015). How quantum theory helps us explain. The British Journal for the Philosophy
of Science, 66, 1–43.
Healey, R. (2017). The quantum revolution in philosophy. Oxford: Oxford University Press.
Heisenberg, W. (1958). Physics and philosophy: The revolution in modern science. London:
George Allen & Unwin.
Howard, D. (1990). Nicht sein kann was nicht sein darf,’ or the prehistory of EPR, 1909–1935:
Einstein’s early worries about the quantum mechanics of composite systems. In A. I. Miller
(Ed.), Sixty-two years of uncertainty: Historical, philosophical, and physical inquiries into the
foundations of quantum mechanics (pp. 61–111). New York/London: Plenum.
Hubert, M., & Romano, D. (2018). The wave-function as a multi-field. European Journal for
Philosophy of Science, 1–17.
Huggett, N., & Wüthrich, C. (2013). Emergent spacetime and empirical (in) coherence. Studies in
Histories and Philosophy of Science, B44, 276–285.
Huggett, N., & C. Wüthrich (Eds.) (forthcoming), Out of nowhere: The emergence of spacetime in
quantum theories of gravity. Oxford: Oxford University Press.
Kitcher, P. (1989). Explanatory unification and the causal structure of the world. In P. Kitcher
& W. Salmon (Eds.), Minnesota studies in the philosophy of science (Vol. 13, pp. 410–503).
Minnesota: University of Minnesota Press.
Knox, E. (2013). Effective Spacetime geometry. Studies in the History and Philosophy of Modern
Physics, 44, 346–356.
Knox, E. (2014). Spacetime structuralism or Spacetime functionalism? Manuscript.
Lam, V., & Esfeld, M. A. (2013). A dilemma for the emergence of Spacetime in canonical quantum
gravity. Studies in History and Philosophy of Modern Physics, 44, 286–293.
Lam, V., & Wüthrich, C. (2017). Spacetime is as Spacetime does. Manuscript.
2 Why Scientific Realists Should Reject the Second Dogma of Quantum Mechanics 47

Leifer, M. S. (2014). Is the quantum state real? An extended review of ψ-ontology theorems.
Quanta, 3, 67–155.
Lewis, P. J. (2004). Life in configuration space. The British Journal for the Philosophy of Science,
55, 713–729.
Lewis, P. J. (2005). Interpreting spontaneous collapse theories. Studies in History and Philosophy
of Modern Physics, 36, 165–180.
Lewis, P. J. (2006). GRW: A case study in quantum ontology. Philosophy Compass, 1, 224–244.
Lewis, P. J. (2013). Dimension and illusion. In D. Albert & A. Ney (Eds.), The wave-function:
Essays in the metaphysics of quantum mechanics (pp. 110–125). New York: Oxford University
Press.
Loewer, B. (1996). Humean Supervenience. Philosophical Topics, 24, 101–127.
Lorentz, H. A. (1909). The theory of electrons. New York: Columbia University Press.
Maudlin, T. (1995). Three measurement problems. Topoi, 14, 7–15.
Maudlin, T. (2007). Completeness, supervenience and ontology. Journal of Physics A, 40, 3151–
3171.
Miller, E. (2014). Quantum entanglement, bohmian mechanics, and humean supervenience.
Australasian Journal of Philosophy, 92(3), 567–583.
Monton, B. (2013). Against 3N-dimensional space. In D. Z. Albert & A. Ney (Eds.), The Wave-
function: Essays in the metaphysics of quantum mechanics (pp. 154–167). New York: Oxford
University Press.
Ney, A. (2012). The status of our ordinary three-dimensions in a quantum universe. Noûs, 46,
525–560.
Ney, A. (2013). Ontological reduction and the wave-function ontology. In D. Albert & A. Ney
(Eds.), The wave-function: Essays in the metaphysics of quantum mechanics (pp. 168–183).
New York: Oxford University Press.
Ney, A. (2015). Fundamental physical ontologies and the constraint of empirical coherence.
Synthese, 192(10), 3105–3124.
Ney, A. (2017). Finding the world in the wave-function: Some strategies for solving the macro-
object problem. Synthese, 1–23.
Ney, A. (forthcoming). Finding the world in the wave function. Oxford: Oxford University Press.
Norsen, T. (2010). The theory of (exclusively) local Beables. Foundations of Physics, 40(12),
1858–1884.
Norsen, T. (2017). Foundations of quantum mechanics: An exploration of the physical meaning of
quantum theory. Cham: Springer.
North, J. (2013). The structure of the quantum world. In D. Z. Albert & A. Ney (Eds.), The wave-
function: Essays in the metaphysics of quantum mechanics (pp. 184–202). New York: Oxford
University Press.
Peierls, R. (1991). In defence of ‘measurement’. Physics World, January, pp. 19–20.
Pitowsky, I. (2007). Quantum mechanics as a theory of probability. In W. Demopoulos & I.
Pitowsky (Eds.), Physical theory and its interpretation: Essays in honor of Jeffrey Bub. Berlin:
Springer.
Przibram, K (1967). Letters on wave mechanics (trans. Martin Klein). New York: Philosophical
Library.
Pusey, M., Barrett, J., & Rudolph, T. (2012). On the reality of the quantum state. Nature Physics,
8, 475–478.
Putnam, H. (1960). Minds and machines, reprinted in Putnam (1975b), pp. 362–385.
Putnam, H. (1975). Mind, language, and reality. Cambridge: Cambridge University Press.
Rovelli, C. (1996). Relational quantum mechanics. International Journal of Theoretical Physics,
35(8), 1637–1678.
Schrödinger, E. (1926). Quantisierung als Eigenwertproblem (Zweite Mitteilung). Annalen der
Physik, 79, 489–527. English translation: Quantisation as a Problem of Proper Values. Part II.
Schrödinger, E. (1935a). Schrödinger to Einstein, 19 August 1935. Translated in Fine, A. (1996).
The Shaky game: Einstein realism and the quantum theory (p. 82). Chicago: University of
Chicago Press.
48 V. Allori

Schrödinger, E. (1935b). Die gegenwärtige Situation in der Quantenmechanik. Die Naturwis-

senschaften, 23, 807–812, 823–828, 844–849.
Sellars, W. (1962). Philosophy and the scientific image of man. In R. Colodny (Ed.), Frontiers of
science and philosophy (pp. 35–78). Pittsburgh, PA: University of Pittsburgh Press.
Spekkens, R. W. (2007). Evidence for the epistemic view of quantum states: A toy theory. Physical
Review A, 75, 032110.
Suárez, M. (2007). Quantum Propensities. Studies in the History and Philosophy of Modern
Physics, 38, 418–438.
Suarez, M. (2015). Bohmian dispositions. Synthese, 192(10), 3203–3228.
Timpson, C. (2010). Rabid dogma? Comments on Bub and Pitowsky. In S. Saunders, J. Bar-
rett, A. Kent, & D. Wallace (Eds.), Many worlds? Everett, quantum theory, and reality
(pp. 460–466). Oxford: Oxford University Press.
Van Camp, W. (2011). On kinematic versus dynamic approaches to special relativity. Philosophy
of Science, 78(5), 1097–1107.
von Neuman, J. (1932). Mathematische Grundlagen der Quantenmechanik. Springer. Translated
by R. T. Beyer as: Mathematical Foundations of Quantum Mechanics. Princeton: Princeton
University, 1955.
Wallace, D. M. (2002). Everettian rationality: Defending Deutsch’s Approach to probability in the
Everett interpretation. Studies in History and Philosophy of Science Part B: Studies in History
and Philosophy of Modern Physics, 34(3), 415–439.
Chapter 3
Unscrambling Subjective and Epistemic
Probabilities

Guido Bacciagaluppi

Abstract There are two notions in the philosophy of probability that are often used
interchangeably: that of subjective probabilities and that of epistemic probabilities.
This paper suggests they should be kept apart. Specifically, it suggests that the dis-
tinction between subjective and objective probabilities refers to what probabilities
are, while the distinction between epistemic and ontic probabilities refers to what
probabilities are about. After arguing that there are bona fide examples of subjective
ontic probabilities and of epistemic objective probabilities, I propose a systematic
way of drawing these distinctions in order to take this into account. In doing so, I
modify Lewis’s notion of chances, and extend his Principal Principle in what I argue
is a very natural way (which in fact makes chances fundamentally conditional). I
conclude with some remarks on time symmetry, on the quantum state, and with
some more general remarks about how this proposal fits into an overall Humean
(but not quite neo-Humean) framework.

Keywords Subjective probabilities · Epistemic probabilities · Principal

principle · Quantum state · Humeanism

I believe I first met Itamar Pitowsky in what I think was the spring of 1993, when he
came to spend a substantial sabbatical period at Wolfson College, Cambridge, while
writing his paper on George Boole’s ‘Conditions of Possible Experience’ (Pitowsky
1994). At that time, the Cambridge group was by far one of the largest research
groups in philosophy of physics worldwide, and Wolfson had the largest share of the
group. Among others, that included two further Israelis, my friends and fellow PhD-
students Meir Hemmo, who is co-editing this volume, and Jossi Berkovitz, who had
worked under Itamar’s supervision on de Finetti’s probabilistic subjectivism and its

G. Bacciagaluppi ()
Descartes Centre for the History and Philosophy of the Sciences and the Humanities, Utrecht
University, Utrecht, Netherlands
e-mail: g.bacciagaluppi@uu.nl

© Springer Nature Switzerland AG 2020 49

M. Hemmo, O. Shenker (eds.), Quantum, Probability, Logic,
Jerusalem Studies in Philosophy and History of Science,
https://doi.org/10.1007/978-3-030-34316-3_3
50 G. Bacciagaluppi

application to quantum mechanics (published in English as Berkovitz 2012). We

were all lunch regulars at Wolfson, and discussions on the philosophy of physics,
of probability, and of mathematics were thus not limited to the setting of the Friday
seminars in the HPS department (or the immediately preceding lunches at the Eraina
Taverna).
My earliest introduction to de Finetti’s subjectivism was in fact through Itamar
and Jossi, and Itamar’s seminal work on quantum probability (Pitowsky 1989, 1994)
looms large within my formative years. The present paper relates most closely
to Itamar’s more recent work on Bayesian quantum probabilities (Pitowsky 2003,
2007). I shall not be agreeing with Itamar on everything, but that fits into the wider
spirit of friendship, co-operation, integrity, and pursuit of knowledge that makes
philosophers of physics an ideal model of an academic community, and Itamar an
ideal model of a philosoper of physics. I dedicate this paper to his memory.

3.1 Subjective or Epistemic Probabilities?

This paper is about the notion of subjective probabilities and that of epistemic
probabilities, and how to keep them apart. Since terminology varies, I should briefly
sketch which uses of the terms I intend. (A fuller explication will be given in
Sects. 3.2 and 3.3.)
I take the use of the term ‘subjective probability’ to be fairly uniform in the
literature, and to be fairly well captured by the idea of degrees of belief (and as such
‘assigned by agents’ and presumably ‘living in our heads’), as opposed to some
appropriate notion of objective probabilities (‘out there’), sometimes cashed out in
terms of frequencies or propensities. I shall take the standard explication of objective
probabilities as that given by David Lewis with his notion of chance (Lewis 1980).
Along these lines, all probability assignments can be taken as subjective in the
first place, but some such assignments may correspond to objective probabilities ‘out
there’ in the world, determined by actual matters of fact. Probability assignments
that do not correspond to objective probabilities shall be termed ‘purely’ or ‘merely’
subjective in the following. In this terminology, the subjective probabilities of a
subjectivist like de Finetti, for whom objective probabilities do not exist, are the
prime example of purely subjective probabilities, even though usually they will
be firmly if pragmatically rooted in matters of actual fact (as further discussed in
Sects. 3.4 and 3.10).
The term ‘epistemic probabilities’ has a wider variety of meanings, of which two
are probably dominant, depending on whether one emphasises the fact that we lack
knowledge of something or whether one emphasises the knowledge that we do have.
The one I shall focus on is the former, that of ignorance-interpretable probabilities
(i.e. of probabilities attaching to propositions expressing a matter of fact, but whose
truth is unknown), which is perhaps most common in the philosophy of physics. The
other refers to probability assignments whose values reflect our state of knowledge.
The classical notion of probability based on the principle of indifference or an
3 Unscrambling Subjective and Epistemic Probabilities 51

objective Bayesian view based on the principle of maximum entropy (to which we
shall presently return) can thus also be termed epistemic.
This second notion of ‘epistemic probabilities’, which is perhaps most common
in the philosophy of probability, belongs somewhere along the axis between subjec-
tive and objective probabilities. It is often contrasted with subjective probabilities
when these are understood as merely ‘prudential’ or ‘pragmatic’ or ‘instrumental’.
This distinction is challenged by Berkovitz (2019), who convincingly argues that de
Finetti’s instrumental probabilities are also epistemic in this sense. Indeed, when
talking about subjective probabilities in Sects. 3.4, 3.6 and 3.10, I shall always
implicitly take them to be so. This second sense of ‘epistemic probabilities’ is not
the topic of this paper, and will be discussed no further.
Epistemic probabilities in the sense of ignorance-interpretable ones instead are
liable to be assimilated to subjective probabilities, but, as I shall argue, they ought
not to, because the distinction between probabilities that are epistemic in this sense
and those that are not (we shall call them ‘ontic’) is conceptually orthogonal to the
subjective-objective distinction.
Take for example Popper’s discussion in Vol. III of the Postscript (Popper
1982).1 After treating extensively of his propensity interpretation in Vol. II, Popper
starts Vol. III (on quantum mechanics) with a discussion of subjective vs objective
probabilities, complaining that physicists conflate the two (especially in quantum
mechanics). Popper’s contrast for the objective view of probabilities is what he
throughout calls the subjective view, which for him is the view that probabilistic
statements are about our (partial) ignorance. In the case of classical statistical
mechanics, he complains that ‘[t]his interpretation leads to the absurd result that the
molecules escape from our bottle because we do not know all about them’ (Popper
1982, p. 109).
True, we care about whether our coffee goes cold, and the coffee does not care
about our feelings. And, at least in the sense in which Popper took his own notion
of propensities to be explanatory of observed frequencies (cf. Bub and Pitowsky
1985, p. 541), it seems reasonable to say that while the behaviour of the coffee is
a major factor in explaining why we come to have certain subjective expectations,
it is only objective probabilities that can explain the behaviour of the coffee. But
there is also a sense in which the probabilities of (classical) statistical mechanics
are clearly epistemic: we do believe the coffee or the gas in the bottle to be in a
definite microstate, but we do not know it. Also the statistical mechanical description
of the gas presupposes there to be a particular matter of fact about the state of
the molecules, which is however left out of the description. If we read Popper’s
complaint as being that epistemic probabilities must be subjective, or rather purely

1 Incidentally, reviewed by Itamar when it came out (Bub and Pitowsky 1985). Note that my
examples of possible assimilation or confusion between epistemic and subjective probabilities (in
the senses just sketched) may well prove to be straw-mannish under further scrutiny, but I just wish
to make a prima facie case that these two notions can indeed be confused. The rest of the paper
will make the extended case for why we should distinguish them and how.
52 G. Bacciagaluppi

subjective (i.e. unrelated to objective probabilities), then it seems that no explanation

of the kind Popper envisages is possible at all in statistical mechanics.
Or take Jaynes, e.g. in ‘Some Random Observations’ (1985). For Jaynes all prob-
abilities are fundamentally subjective in the sense that ‘a probability is something
that we assign in order to represent a state of knowledge’, as opposed to an actual
frequency, which is a ‘factual property of the real world’ (p. 120). While for Jaynes
the probabilities of classical statistical mechanics are thus of course epistemic,
unlike for Popper they are meant to be (at least partially?) objective, in the sense
that there are (rational?) principles, first and foremost the principle of maximum
entropy, that provide a way for choosing our probability assignments based on those
facts we do know.2
Part of Jaynes’s motivation came from the complaint that the mathematics of
quantum theory ‘describes in part physical law, in part the process of human
inference, all scrambled together in such a way that nobody has seen how to separate
them’ (p. 123). He continues:
Many years ago I became convinced that this unscrambling would require that probability
theory itself be reformulated so that it recognizes explicitly the role of human information
and thus restores the distinction between reality and our knowledge of reality, that has been
lost in present quantum theory.3

Jaynes’ position, however, faces a different problem precisely in the case of

quantum mechanics. Whether or not probabilities are purely subjective, for Jaynes
they are always epistemic: ‘[p]robabilities in present quantum theory express the
incompleteness of human knowledge just as truly as did those in classical statistical
mechanics’ (1985, p. 122). And this ominously leads him to question the possibility
of fundamental indeterminism.

2 It is unclear to what extent this ‘objective Bayesian’ strategy indeed yields objective probabilities.

Consider for instance situations in which information about the physical situation is in principle
readily available but is glibly ignored. In this case, even if we use principles such as maximum
entropy, one might complain that there is little or no connection between the physical situation
and the probabilities we assign. On the other hand, there may be situations in which the
physical situation itself appears to systematically lead to situations of ignorance, e.g. because of
something like mixing properties of the dynamics. In this case, there may arguably be a systematic
link between certain objective physical situations and certain ways of assigning probabilities,
which may go some way towards justifying the claim that these probabilities are objective. (Cf.
discussions about whether high-level probabilities are objective, e.g. Glynn (2009), Lyon (2011),
Emery (2013) and references therein.) If so, note that this will be the case whether or not there
is anyone to assign those probabilities. Many thanks to Jossi Berkovitz for discussion of related
points.
3 For an example of what Jaynes has in mind, see Heisenberg’s ‘The Copenhagen interpretation

of quantum theory’ (Heisenberg 1958, Chap. 3). In Sect. 3.9 below, I shall allude to what I think
are the origins of these aspects of Heisenberg’s views, but I agree with Jaynes that (at least in the
form they are usually presented) they are hard to make sense of. Note also that the ‘Copenhagen
interpretation’ as publicised there appears to have been largely a (very successful) public relations
exercise by Heisenberg (cf. Howard 2004). In reality, physicists like Bohr, Heisenberg, Born and
Pauli held related but distinct views about how quantum mechanics should be understood.
3 Unscrambling Subjective and Epistemic Probabilities 53

Indeed, as Jaynes sees it, in the case of quantum mechanics human knowledge is
limited by dogmatic adherence to indeterminism in the Copenhagen interpretation.
But in fact, even though the ‘Copenhagen interpretation’ may have been dogmatic,
indeterminism was the least worrying of its dogmas: a wide spectrum of non-
dogmatic approaches to the foundations of quantum mechanics today embrace
indeterminism in some form or other. Thus, if (as Jaynes seems to think) subjective
probabilities must be epistemic, then they seem to be inappicable to many if not
most fundamental approaches to quantum mechanics.4
These problems with both Popper’s and Jaynes’ positions seem to me to stem
from a scrambling together of subjective and epistemic probabilities. I shall take it
as uncontroversial that we often form degrees of belief about facts we do not know
(so that such probabilities are both subjective and epistemic), but this in itself does
not show that all epistemic probabilities should be subjective, or that all subjective
probabilities should be epistemic. Indeed, being a degree of belief is conceptually
distinct from being a probability about something we ignore: the characterisation of
a probability as subjective is in terms of what that probability is, that of a probability
as epistemic is in terms of what that probability is about.
Admittedly, even if we can conceptually distinguish between subjective and
epistemic probabilities, it could be that the two categories are coextensive. And
if all epistemic probabilities are merely subjective, or all subjective probabilities
are also epistemic, then Popper’s or Jaynes’s worries become legitimate again. As
a matter of fact, typical cases of epistemic probabilities are arguably subjective (‘Is
Beethoven’s birthday the 16th of December?’5 ), and, say, probabilities in plausibly
indeterministic scenarios are typically thought to be objective (‘Will this atom decay
within the next minute?’). And, indeed, while there is an entrenched terminology
for probabilities that are not purely subjective (namely ‘objective probabilities’),
non-epistemic probabilities do not even seem to have a separate standard name (we
shall call them ‘ontic probabilities’). It is indeed tempting to think that all epistemic
probabilties are subjective and all ontic probabilities are objective.
I urge the reader to resist this temptation. I believe that we can make good
use of the logical space that lies between these distinctions. As the Popper and
Jaynes quotations above are meant to suggest, objectivism about probability will be

4 Spontaneous collapse approaches are invariably indeterministic (unless one counts among them
also non-linear modifications of the Schrödinger equation, which however have not proved
successful), as is Nelson’s stochastic mechanics. Everett’s theory vindicates indeterminism at
the emergent level of worlds. And even many pilot-wave theories are stochastic, namely those
modelled on Bell’s (1984) theory based on fermion number density. Thus, among the main
fundamental approaches to quantum mechanics, only theories modelled on de Broglie and Bohm’s
original pilot-wave theory provide fully deterministic underpinnings for it. (For a handy reference
to philosophical issues in quantum mechanics, see Myrvold (2018b) and references therein. For
more on Nelson see Footnote 40.)
5 16 December is the traditional date for Beethoven’s birthday, but we know from documentary

evidence only that he was baptised on 17 December 1770 – Beethoven was probably born on 16
December.
54 G. Bacciagaluppi

self-limiting if it takes itself to apply only to ontic probabilities, as will subjectivism

about probability if it takes itself to apply only to epistemic probabilities. The
landscape of the philosophy of probability will be both richer and more easily
surveyable if we unscramble subjective and epistemic probabilities in a way that
allows also for (purely) subjective ontic probabilities as well as for epistemic
objective probabilities. While this is surely not the first paper to make some of these
claims (indeed, Jaynes is obviously proposing that epistemic probabilities can be
objective6 ), I believe the way it approaches these questions will be original.
In what follows, I shall first recall the standard discussion of subjective and
objective probabilities given by David Lewis (Sect. 3.2), then try and spell out what
one typically means by the distinction between epistemic and ontic probabilities
(Sect. 3.3). This will provide enough background to argue that there are bona fide
cases of probabilities that are both purely subjective and ontic (Sect. 3.4), and of
probabilities that are both epistemic and objective (Sect. 3.5). This will lead to
an extension of Lewis’s definition of objective probabilities and of his Principal
Principle in which chances become fundamentally conditional, thus redrawing the
subjective-objective distinction (Sect. 3.6), and to a more systematic discussion also
of the epistemic-ontic distinction, clarifying some further ambiguity (Sect. 3.7). I
conclude the paper with some additional remarks on time symmetry and on the
quantum state (Sects. 3.8 and 3.9),7 and with more general remarks on how the
proposed framework for conceptualising probabilities fits into an overall Humean
framework – at the same time distancing myself from some typical neo-Humean
views (Sect. 3.10).

3.2 Subjective vs Objective Probabilities

Recall the standard distinction between subjective and objective probabilities:

subjective probabilities are coherent degrees of belief, generally taken to live in our
head, while objective probabilities are meant to live out there in the world, properties
of external objects or reduced to such properties. I shall follow (half of) the literature
in taking subjective probabilities as well-understood, and objective probabilities (if
they exist) as being the more mysterious concept. While the literature on subjective
probabilities is considerable and varied, I shall follow what is nowadays often
considered the standard analysis of how subjective and objective probabilities relate
– and which in fact provides an implicit definition of objective probabilities given an

6 With the qualifications mentioned in Footnote 2. Thanks to both Jos Uffink and Ronnie Hermens
for making me include more about Jaynes.
7 These (especially the latter) can be skipped by readers who are not particularly familiar with

debates on the metaphysics of time or, respectively, on the nature of the quantum state and of
collapse in quantum mechanics.
3 Unscrambling Subjective and Epistemic Probabilities 55

understanding of subjective probabilities – David Lewis’s ‘A Subjectivist’s Guide to

Objective Chance’ (Lewis 1980).8
Importantly for our later discussion, Lewis takes objective probabilities (also
called ‘chances’) to be functions of propositions and times. Assuming for the sake
of example that there are chances for rolls of dice, before I roll a pair of dice we
have the intuition that the chance of rolling 11 is 1/18. Once I have rolled the first
die, the chance of 11 either goes down to 0 and stays there (if I have rolled one of
the numbers 1 to 4), or it goes up to 1/6 (if I have rolled 5 or 6). Then when I roll
the second die the chance of 11 goes definitively either down to 0 or up to 1.
The idea behind Lewis’s analysis is that objective probabilities, if they exist, are
the kind of thing that – in a sense to be made precise – would rationally constrain
our subjective probabilities (also called ‘credences’9 ) completely. Credences of
course are standardly taken to be rationally constrained both synchronically by the
probability calculus and diachronically by Bayesian conditionalisation.10 Few other
requirements are widely recognised as able to constrain their values further (maybe
only the requirement that credences should be non-dogmatic, i. e. not take the values
0 or 1, is one commanding near-universal support – but see the remarks below in
Sect. 3.4). Thus, the criterion that Lewis proposes for objective probabilities is very
stringent.
What Lewis has in mind is the following. The chance at t of A is x iff there
is a proposition cht (A) = x such that (subject to a few provisos to be spelled
out) conditionalising on cht (A) = x will always yield the unique value x for our
credence,
  

crt A  cht (A) = x = x, (3.1)

whatever the form of our credence function at the time t. Of course, if we further
believe at t that the chance at t of A is in fact x, i. e. if crt (cht (A) = x) = 1, then
our credence at t in A is also x (by the probability calculus), or if we learn at t that
cht (A) = x, our credence in A becomes x (by Bayesian conditionalisation).
The stringency of Lewis’s requirements on what is to count as rationally
compelling is such that there are few candidates in the literature for satisfying them,
none of which are uncontroversial. And because Lewis himself believes in ‘Humean

8 While in a Cambridge HPS mood, I cannot refrain from thanking Jeremy Butterfield for (among
so many other things!) introducing me to this wonderful paper and all issues around objective
probabilities.
9 Again, terminology varies somewhat, but I shall take ‘credences’ to be synonymous with

‘subjective probabilities’.
10 Bayesian conditionalisation is often taken as the unique rational way in which one can change

one’s credences. But there may be other considerations competing with it, in particular affecting
background theoretical knowledge. This will generally be implicit (but not needed) in much of this
paper. Thanks to Jossi Berkovitz for pointing out to me that de Finetti himself was not committed
to conditionalisation as the only way for updating probabilities, believing instead that agents may
have reasons for otherwise changing their minds (cf. Berkovitz 2012, p. 16).
56 G. Bacciagaluppi

supervenience’, for him chances have to supervene on ‘particular matters of fact’,

making it especially hard to find plausible explicit candidates for objective chances.
This problem is quaintly known in the literature as the ‘big bad bug’.11
The provisos are, first, that our credence function at t should not rule out that
cht (A) = x, i. e. we should have crt (cht (A) = x) = 0, otherwise the credence in
(3.1) will be ill-defined; second, and more subtly, that our credence function at t
should not incorporate information that is ‘inadmissible’ at t. To make this intuitive,
think of the chance at t of A as fixing the objectively fair odds for betting on A.
Clearly, if in advance of making the bet either you or the bookie had some more
direct information about whether or not A will come to pass (e. g. one of you is
clairvoyant), that would be cheating. In order for the notion of ‘objectively fair
odds’ at time t to make sense, we need to have an appropriate notion of what is
admissible in shaping our credences at t. Not everything can be admissible, for
then we would allow A, or not-A, and the only propositions about chance that
could satisfy (3.1) would correspondingly be propositions implying cht (A) = 1,
or propositions implying cht (A) = 0, and all chances would be trivial. On the other
hand, some propositions must be admissible, since cht (A) = x clearly is. We can
write our credence function crt (A) as cr(A|Et & T ), where cr(A) is a ‘reasonable
initial credence function’, T represents any background assumptions that we make
(which might perhaps better go in a subscript), and Et is our epistemic context at
time t (by which I mean all propositions about matters of fact known at time t). If
we leave fixed the background assumptions T , then crt (A) just evolves by Bayesian
conditionalisation upon the new facts that we learn between any two times t and
t  .12

11 Two proposals of note are the one by Deutsch and by Wallace in the context of Everettian
quantum mechanics, who derive the quantum mechanical Born rule from what they argue are
rationality constraints on decision making in the context of Everett (see e.g. Wallace 2010, or
Brown and Ben Porath (2020)), and the notion of ‘Humean objective chances’ along the lines of
the neo-Humean ‘best systems’ approach (Loewer 2004; Frigg and Hoefer 2010). Note that I am
explicitly distinguishing between Lewisian chances and Humean objective chances. The latter are
an attempt to find something that will satisfy the definition of Lewisian chances, but there is no
consensus as to whether it is a successful attempt. (See Sect. 3.10 below for some further comments
on neo-Humeanism.) The Deutsch-Wallace approach is often claimed to provide the only known
example of Lewisian chances, but again this claim is controversial, as it appears that some of the
constraints that are used cannot be thought of as rationality assumptions but merely as naturalness
assumptions. On this issue see Brown and Ben Porath (2020) as well as the further discussion in
Saunders et al. (2010). The difficulty in finding anything that satisfies the definition of Lewisian
chances is (of course) an argument in favour of subjectivism.
12 I shall assume throughout that degrees of belief attach to propositions about matters of fact (in

the actual world), which incidentally need not be restricted to Lewis’s ‘particular’ matters of fact
(e.g. I shall count as matters of fact also holistic facts about quantum states – if such facts there
be). I shall call these material or empirical propositions.
Of course, some of the background assumptions that we make in choosing our degrees of belief
may not just refer to matters of fact, but will be theoretical propositions (say, referring to possible
worlds). In that case, I would prefer not to talk about degrees of belief attaching to such theoretical
propositions, but about degrees of acceptance.
3 Unscrambling Subjective and Epistemic Probabilities 57

Substituting into (3.1), we have

  

cr A  (cht (A) = x) & Et & T = x. (3.2)

And the requirement that (3.1) should hold whatever the form of our credence
function at t (subject to the provisos) now translates into the requirement that (3.2)
should hold for any ‘reasonable initial credence function’ (which we shall keep
fixed) and for any admissible propositions Et and T , which for Lewis are, indeed,
ones that are compatible with the chance of A being x and in addition do not provide
information about the occurrence of A over and above what is provided by knowing
the chance of A (if chance is to guide credence, we must not allow information that
trumps the chances). This is Lewis’s ‘Principal Principle’ (PP) relating credence and
chance. It can be taken, is often taken, and I shall take it as providing the implicit
definition of objective probabilities: iff there is a proposition X such that, for all
choices of reasonable initial credence function and admissible propositions,
  

cr A  X & Et & T = x, (3.3)

The matter is further complicated by the fact that even full acceptance of a theory need not
imply belief in all propositions about matters of fact following from the theory. For instance, an
empiricist may believe only a subset of the propositions about actual matters of fact following
from a theory they accept (say, only propositions about observable events), thus drawing the
empirical-theoretical distinction differently and making a distinction between material propositions
and empirical propositions, which for the purposes of this paper I shall ignore. (For my own very
real empiricist leanings, see Bacciagaluppi 2019.)
For simplicity I shall also ignore the distinction between material propositions and actual
propositions that one makes in a many-worlds theory, where the material world consists of several
‘possible’ branches, even though I shall occasionally refer to Everett for illustration of some point.
One might reflect the distinction between material (or empirical) and theoretical propositions
by not conditionalising on the latter, but letting them parameterise our credence functions. Another
way out of these complications may be to assign subjective probabilities to all propositions alike,
but interpret the probabilities differently depending on what propositions they attach to: in general
they will be degrees of acceptance, but in some cases (all material propositions, or a more restricted
class of empirical propositions), one may think of them as straightforward degrees of belief. Such
an additional layer of interpretation of subjective probabilities will, however, not affect the role
they play in guiding our actions (degrees of acceptance will generally be degrees of ‘as-if-belief’).
In any case, the stroke notation makes clear that assuming a theory – even hypothetically – is meant
to affect our credences as if we were conditionalising on a proposition stating the theory (I believe
this is needed to derive some of Lewis’s results).
Note that for Lewis, T will typically include theoretical propositions about chances themselves
in the form of history-to-chance conditionals, and because in Lewis’s doctrine of Humean
supervenience chances in fact supervene on matters of fact in the actual world, such background
assumptions may well be propositions that degrees of belief can attach to (at least depending on
how one interprets the conditionals), even if one wishes to restrict the latter (as I have just sketched)
to propositions about matters of fact in the actual world.
We shall return to the material-theoretical distinction in Sect. 3.7.
58 G. Bacciagaluppi

then there is such a thing as the objective chance at t of A and its value is x. The
weakest such proposition X is then cht (A) = x.13
More precisely, Lewis restricts admissible propositions about matters of fact to
historical propositions up to time t (at least ‘as a rule’: no clairvoyance, time travel
and such, and we shall follow suit for now – although he appears to leave it open that
one may rethink the analysis if one should wish to consider such less-than-standard
situations (Lewis 1980, p. 274)). Admissible propositions about matters of fact can
thus be taken to be what can be reasonably thought to be knowable in principle at
t, and Lewis allows them to include the complete history Ht of the world up to that
time. He also allows hypothetical propositions about the theory of chance, i. e. about
how chances may depend on historical facts.
Even more propositions might be admissible at time t, but this partial char-
acterisation of admissibility already makes the PP a useful and precise enough
characterisation of objective chances. Indeed, taking the conjunction Ht & TC of
the complete history of the world up to t and a complete and correct theory TC of
how chances depend on history, the following will be an instance of the PP:
  

cr A  (cht (A) = x) & Ht & TC = x. (3.4)

Since TC is a complete and correct theory of chance, Ht & TC already implies that
cht (A) = x, and we can rewrite (3.4) as

cht (A) = cr(A|Ht & TC ). (3.5)

This instance of the PP is used by Lewis to derive various properties of chances

using known properties of credences, e. g. that they satisfy the probability calculus,
and that they evolve by conditionalisation on intervening events. In particular,
according to Lewis, chances of past events are all 0 or 1. Taking ch(A) as the
chance function at some arbitrary time t = 0, we can thus for all later times write
cht (A) = ch(A|Ht ).

3.3 Epistemic vs Ontic Probabilties

A systematic treatment of the epistemic-ontic distinction comparable to Lewis’s

treatement of subjective and objective probabilities is not available. But we can get
a reasonably good idea of such a distinction, which will be enough for the time
being, by looking at common ways of using the term ‘epistemic probabilities’. (We
shall revisit this distinction in Sect. 3.7.)
Epistemic probabilities are probabilities that attach to propositions of which we
are ignorant. An obvious example is: I flip a coin and do not look at the result

13 Notethat X is a proposition about (actual) matters of fact. Thus, chances in the sense of Lewis
indeed satisfy the intuition that objective probabilities are properties of objects or more generally
determined by actual facts. Thanks to Jossi Berkovitz for discussion.
3 Unscrambling Subjective and Epistemic Probabilities 59

yet – what is the probability of Heads? In this situation, there is a matter of fact
about whether the coin has landed Heads or Tails, and if I knew it, I would assign to
Heads probability 0 or 1 (depending on what that matter of fact was).
Another common example is modelling coin flipping as a deterministic Newto-
nian process: even before the coin has landed, the details of the flip (initial height,
velocity in the vertical direction, and angular momentum14 ) determine what the
outcome will be, but we generally do not know these details. In this case, we say
that the probabilities of Heads and Tails are epistemic, even if the event of the coin
landing Heads or Tails is in the future, because we can understand such events as
labels for sets of unknown initial conditions.
If we are looking for examples of ontic probabilities, we thus presumably
need to look at cases of indeterminism. The most plausible example is that of
quantum probabilities, say, the probability for a given atom to decay within the next
minute. (Of course, whether this is a genuine case of indeterminism will depend on
one’s views about quantum mechanics, in particular about the existence of ‘hidden
variables’. For the sake of the example we shall take quantum mechanics to be
genuinely indeterministic.)
Still, if all actual matters of fact past, present and future are taken to be knowable
in principle, then all probabilities become purely epistemic, and (at least for now) it
appears we cannot draw an interesting distinction.
One way of ensuring that the future is unknowable, of course, is to postulate that
the future is open, i. e. that at the present moment there is no (single) future yet.
This is a very thick metaphysical reading of indeterminism, however, which I wish
to avoid for the purpose of drawing the epistemic-ontic distinction.15 Indeterminism
can be defined in a non-metaphysical way in terms of the allowable models of
a theory (or nomologically possible worlds, if worlds are also understood in a
metaphysically thin way): a theory is (future-)deterministic iff any two models
agreeing up to (or at) time t agree also for all future times (and with suitable
modifications for relativistic theories).
For the purpose of this first pass at drawing the epistemic-ontic distinction we
can then simply stipulate that the notion of ignorance only pertains to facts in the
actual world that are knowable in principle, and that at time t only facts up to t
are knowable in principle (which is in fact the stipulation we already made in the
context of Lewis’s analysis of chances – we may want again to make a mental note
to adapt the notion of ‘knowable in principle’ if we allow for clairvoyance, time

14 For the classic Newtonian analysis of coin flipping see Diaconis et al. 2007.
15 Or for any other purpose – as a matter of fact, I shall assume throughout a block universe picture.

Specifically, in the case of indeterminism one should think of each possible world as a block
universe. The actual world may be more real than the merely possible ones, or all possible worlds
may be equally real (whether they are or not is inessential to the following, although a number
of remarks below may make it clear that I am no modal realist, nor in fact believe in objective
modality). But within each world, all events will be equally real. (Note that at a fundamental level
the Everett theory is also deterministic and Everettians emphasise that the Everettian multiverse is
indeed a block universe.)
60 G. Bacciagaluppi

travel and such like). If one so wishes, one can of course talk about ‘ignorance’ of
which possible world is the actual world, or of which way events in our world will
be actualised. But if probabilities in cases of indeterminism are to count as ‘ontic’,
then it simply is the case that in practice we make a more restrictive use of the term
‘ignorance’ when we draw the epistemic-ontic distinction.
The general intuition is now that epistemic probabilities lie in the gap between
probability assignments given what is knowable in principle (call it the ontic
context), and probability assignments given what is actually known (the epistemic
context). By the same token, ontic probabilities are any probabilities that remain
non-trivial if we assume that we know everything that is knowable in principle.
In general, probabilities will be a mix of epistemic and ontic probabilities. For
instance, we ask for the probability that any atom in a given radioactive probe
may decay in the next minute, but the composition of the probe is unknown:
the probability of decay will be given by the ontic probabilities corresponding
to each possible radioactive component, weighted by the epistemic probabilities
corresponding to each component.
The EPR setup in quantum mechanics (within a standard collapse picture)
provides another good example: before Bob performs a spin measurement, assuming
the global state is the singlet state, Alice has ontic probability 0.5 of getting spin up
in any direction. After Bob performs his measurement, Alice has ontic probabilities
cos2 (α/2) and sin2 (α/2) of getting up or down, where α is the angle between the
directions measured on the two sides. However, she is ignorant of Bob’s outcome,
and her epistemic probabilities for his having obtained up or down are both 0.5.
Thus, given her epistemic context, her (now mixed) probabilities for her own
outcomes are unchanged.16
As a final example, assume we have some ontic probability that is also an
objective chance in the sense of Lewis. According to the PP in the form (3.5), this
ontic probability will be given by

cht (A) = cr(A|Ht & TC ). (3.6)

Now compare it with cr(A|Et & TC ), our credence in A at t given the same
theoretical assumptions but our actual epistemic context Et . This now is a mixed
(partially epistemic) probability. Indeed, let Hti be the various epistemic possibilities
(compatible with our epistemic context Et ) for the actual history up to t (i. e. for the
actual ontic context). We then have

cr(A|Et & TC ) = cr(A|Hti & Et & TC )cr(Hti |Et & TC ), (3.7)
i

16 This is of course the quantum mechanical no-signalling theorem, as seen in the standard collapse

picture. (It is instructive to think through what the theorem means in pilot-wave theory or in Everett.
Thanks to Harvey Brown for enjoyable discussion of this point.) And of course, mixed quantum
states are the classic example of the distinction in quantum mechanics between probabilities that
are ignorance-interpretable (‘proper mixtures’) and probabilities that are not (‘improper mixtures’).
3 Unscrambling Subjective and Epistemic Probabilities 61

which simplifies to

cr(A|Et & TC ) = cr(A|Hti & TC )cr(Hti |Et & TC ), (3.8)
i

because the Hti are fuller specifications of the past history at t than the one
provided by Et . In (3.8), now, the credences cr(Hti |Et & TC ) are purely epistemic
probabilities, and the cr(A|Hti & TC ) are purely ontic probabilities corresponding to
each of the epistemically possible ontic contexts. The various Hti could e. g. fix the
composition of the radioactive probe in our informal example above.17
Judging by the examples so far, the subjective-objective distinction and the
epistemic-ontic distinction could well be co-extensive. Indeed, both subjective and
epistemic probabilities seem to have to do with us, while both objective and ontic
probabilities seem to have to do with the world. Epistemic probabilities depend
on epistemic contexts, as do subjective ones; and chances depend on past history,
which we have also taken as the context in which ontic probabilities are defined. But
appearances may deceive.
In order to argue that the subjective-objective distinction and the epistemic-ontic
distinction are to be drawn truly independently, I need to convince you that there
are bona fide examples of probabilities that are both subjective (in fact purely
subjective) and ontic, and of probabilities that are both epistemic and objective.
These will be the respective objectives of the next two sections.

3.4 Subjective Ontic Probabilities

Let us start with the case of subjective ontic probabilities. I have discussed this case
before, in a paper (Bacciagaluppi 2014) in which I comment on Chris Fuchs and co-
workers’ radical subjectivist approach to quantum mechanics, known as qBism (see
e.g. Fuchs 2002, 2014),18 and in which I spell out how I believe radical subjectivism
à la de Finetti (1970) can be applied to quantum mechanics not only along qBist
lines but also within any of the traditional foundational approaches (Bohm, GRW
and Everett). The idea that subjectivism is applicable to quantum mechanics had

17 In this example, it is of course possible to think of mixed epistemic and ontic probabilities as
arising not only through ignorance of matters of fact in the ontic context Ht , but also through
ignorance of the correct theory of chance TC . We might indeed know the exact composition of
the radioactive probe, but take ourselves to be ignorant of the correct decay laws for the various
isotopes, with different credences about which laws are correct. I shall neglect this possibility in
the following, because as noted already in Footnote 12, I prefer not to think of ignorance when
talking about theoretical assumptions. In any case, I shall not need this further possible case for the
arguments below.
18 I am deeply grateful to Chris Fuchs for many lovely discussions about qBism and subjectivism

over a number of years.

62 G. Bacciagaluppi

been already worked out by Jossi Berkovitz in the years 1989–1991 (prompted
by Itamar), with specific reference to de Finetti’s idea that probabilities are only
definable for verifiable events.
For the purposes of the present paper, I need to make the case that purely
subjective probabilities can be ontic, thus radical subjectivism about quantum
probabilities (in a context in which quantum mechanics is taken to be irreducibly
indeterministic) provides the ideal example, and in the following I shall rely heavily
on my own previous discussion. (Alternatively, I could arguably have relied on
Berkovitz 2012.) Note that what I need to convince you of is not that such a position
is correct, but only that it is coherent, i. e. that there is an interesting sense in which
one can combine the notion of purely subjective probabilities with that of ontic
probabilities.
Itamar himself (Pitowsky 2003, 2007), Bub (2007, 2010) and Fuchs and co-
workers have all defended approaches to quantum mechanics in which probabilities
are subjective in the sense of being degrees of belief, and some modifications or
additions to Bayesian rationality are introduced in order to deal with situations,
respectively, in which one deals with ‘quantum bets’ (incompatible bets possibly
with some outcomes in common), in which one assumes no-cloning as a funda-
mental principle, or in which gathering evidence means irreducibly intervening into
the world. I take it that all of these authors (as well as de Finetti himself) agree on
the assumption that in general there are no present or past facts that will determine
the outcomes of quantum bets or any quantum measurements, and thus that the
probabilities in their approaches are ontic in the sense used in the previous section.
It is less clear that the probabilities involved are purely subjective. Indeed, both
Pitowsky and Bub see subjective probability assignments as severely constrained
by the structure of quantum bets or quantum observables, in particular through
Gleason’s theorem and (relatedly) through the mathematical structure of collapse.
These might be seen as rationality constraints, so that the probabilities in these
approaches might in fact be seen as at least partly objective.19 As Itamar puts it
(Pitowsky 2003, p. 408):

19 Whether or not they are will make little difference to my argument (although if they should be
purely subjective I would have further concrete examples of subjective ontic probabilities in the
sense I need). In Bub’s case, he explicitly states that the no-cloning principle is an assumption
and could turn out to be false. That means that the constraints hold only given some specific
theoretical assumptions, and are not rationally compelling in the strong sense used by Lewis. In
Pitowsky’s case, the constraints seem stronger, in the sense that the ‘quantum bets’ are not per se
theoretical assumptions about the world, but are just betting situations whose logical structure is
non-classical. Whether or not there are rational constraints on our probability assignments will be
a well-posed question only after you and the bookie have agreed on the logical structure of the bet.
On the other hand, whenever we want to evaluate credences in a particular situation, we need to
make a theoretical judgement as to what the logic of that situation is. That means, I believe, that
whether we can have rationally compelling reasons to set our credences in any particular situation
will still depend on theoretical assumptions, and fall short of Lewis’s very stringent criterion.
See Bacciagaluppi (2016) for further discussion of identifying events in the context of quantum
probability, and Sect. 1.2 of Pitowsky (2003) for Itamar’s own view of this issue. Cf. also Brown
and Ben Porath (2020).
3 Unscrambling Subjective and Epistemic Probabilities 63

For a given single physical system, Gleason’s theorem dictates that all agents share a
common prior probability distribution or, in the worst case, they start using the same
probability distribution after a single (maximal) measurement.

Indeed, since pure quantum states always assign probability 1 to results of some
particular measurement, the constraints may be so strong that even some dogmatic
credences are forced upon us.
Fuchs instead clearly wishes to think of his probabilities as radically subjective
(which I want to exploit). Precisely because of the strong constraints imposed by
collapse, he feels compelled to not only take probability assignments as subjective,
but also quantum states themselves and even Hamiltonians (Fuchs 2002, Sect. 7).
To a certain extent, however, I think this is throwing out the baby without getting
rid of the bath water. Indeed, I believe the (empirically motivated but normative)
additions to Bayesian rationality to which Fuchs is committed imply already by
themselves very strong constraints on qBist probability assignments (Bacciagaluppi
2014, Sect. 2.2). On the other hand, as I shall presently explain, I also believe that
radical subjectivism can be maintained even within a context in which quantum
states and Hamiltonians are taken as (theoretical) elements in our ontology. While
in some ways less radical than Fuchs’s own qBism, this is the position that I propose
in my paper, and which I take to prove by example that a coherent case can be made
for (purely) subjective ontic probabilities. It can be summarised as follows.
Subjectivism about probabilities (de Finetti 1970) holds that there are no
objective probabilities, only subjective ones. Subjective probability assignments are
strategies we use for navigating the world. Some of these strategies may turn out to
be more useful than others, but that makes them no less subjective. As an example,
take again (classical) coin flipping: we use subjective probabilities to guide our
betting behaviour, and these probabilities get updated as we go along (i.e. based on
past performance). But it is only a contingent matter that we should be successful at
all in predicting frequencies of outcomes. Indeed, any given sequence of outcomes
(with any given frequency of Heads and Tails) will result from appropriate initial
conditions.
Under certain assumptions the judgements of different agents will converge
with subjective probability 1 when conditionalised on sufficient common evidence,
thus establishing an intersubjective connection between degrees of belief and
frequencies. This is de Finetti’s theorem (see e.g. Fuchs 2002, Sect. 9). One
might be tempted to take it as a sign that subjective probabilities are tracking
something ‘out there’.20 This, however, does not follow – because the assumption
of ‘exchangeability’ under which the theorem can be proved is itself a subjective
feature of our probability assignments. The only objective aspect relevant to the
theorem is the common evidence, which is again just past performance.
For the subjectivist, there is no sense in which our probability judgements are
right or wrong. Or as de Finetti (1970, Preface, quoted in Fuchs 2014) famously

20 Such a reading is e.g. perhaps present in Greaves and Myrvold (2010), or even in Pitowsky
(2003).
64 G. Bacciagaluppi

expressed it: ‘PROBABILITIES DO NOT EXIST’. While this sounds radical, it

is indeed perfectly unremarkable as long as we assume that states of affairs are
always determinate and evolve deterministically. In that case, Lewis himself will
say that objective chances are all 0 or 1, so that all non-trivial probabilities are
indeed subjective and (non-trivial) chances do not exist.
What we need to consider is the case of indeterminism. Surely the situation is
completely different in quantum mechanics? As Lewis puts it in the introduction to
the ‘Subjectivist’s Guide’, ‘the practice and analysis of science require both [sub-
jective and objective probabilities]’. He accordingly demands that also subjectivists
should be able to make sense of the proposition that ‘any tritium atom that now
exists has a certain chance of decaying within a year’ (Lewis 1980, p. 263). Thus,
we are told, de Finetti’s eliminative strategy towards chance cannot be applied to
quantum mechanics.
But, actually, why not? In the deterministic case, de Finetti’s analysis shows us
how we can have a useful and even intersubjectively robust notion of probability
that relates to actual matters of fact but is conceptually quite autonomous from them.
And after all, it is especially when we take determinism to fail and the world around
us to behave randomly that we may most need (probabilistic) strategies to navigate
the world.
What goes unappreciated in the sweeping intuition that subjectivism must fail
in the case of indeterminism is precisely the role and status of the strategies we
apply in choosing our probability assignments. Subjectivists may well use certain
strategies or shortcuts to decide what subjective probability assignments to make.
These are just what we called theoretical assumptions or models in Sect. 3.2. They
are not Lewisian rationality constraints on probability assumptions, but systematic
ways in which we subjectively choose our credences. Data may well ‘confirm’ such
models, but again that just means that our subjective probability assignments have
performed well in the past.
The form theoretical models may typically take is that they will systemati-
cally connect certain non-probabilistic features of a given situation with certain
probability assignments. For instance, we may choose to assign certain subjective
probabilities to rolls of dice given their internal weight distribution. For the
subjectivist, these probabilities are and remain purely subjective. Indeed, there is
no necessary connection between the weight distribution in a die and the outcomes
of our rolls, not just in the Humean sense that these are independent matters of fact,
but even in the sense that, given a Newtonian model of dice rolling, for any possible
sequence of results there are initial conditions that will produce it. Thus, there is no
Lewisian compelling rational justification for letting our credences depend on the
weight distribution in the die, only pragmatic criteria such as simplicity, again past
performance, etc.
From this perspective, quantum mechanics is just such a theoretical model for
choosing our subjective probabilities, and indeed a very successful one. One can
take the analogy with probabilistic models of dice rolling quite far: unlike Fuchs,
who identifies quantum states with catalogues of subjective probabilities, I believe a
subjectivist may take the quantum state itself as a non-probabilistic, ontic feature of
3 Unscrambling Subjective and Epistemic Probabilities 65

the world and use it (like weight distribution) as the basis for choosing our subjective
probabilities.21 There is no need for a qBist like Fuchs to reject the putative onticity
of quantum states,22 precisely because that by itself does not make our probability
assignments objective (forgive me, Itamar!).
As further described in Bacciagaluppi (2014), this general perspective can be
applied to all the main approaches to quantum mechanics that take the quantum
state as ontic, so that one can be a de Finettian subjectivist about probabilities in
the context of Bohm, GRW or Everett. In the case of Bohm, this is unremarkable of
course, because the Bohm theory is deterministic and probabilities are epistemic
in the sense used above. In the case of GRW, it is an unfamiliar position. The
default intuition in GRW is presumably to take probabilities as some kind of
propensities, but subjectivism may prove more attractive.23 Finally, the Deutsch-
Wallace decision-theoretic framework for the Everett theory already analyses
probabilities as subjective. Deutsch and Wallace themselves (see e.g. Wallace 2010)
argue that the weights in the universal wavefunction in fact satisfy the Principal
Principle (and are in fact the only concrete example of something that qualifies
as objective chances in Lewis’s sense!). Many authors, however, dispute that the
Deutsch-Wallace theorem rationally compels one to adopt the Born rule. I concur,
but see that as no reason to reject the decision-theoretic framework itself. One can
simply use the Deutsch-Wallace theorem as an ingredient in our subjective strategy
for assigning probabilities on the basis of quantum states. I take it that this is in fact
the subjectivist (or ‘pragmatist’) position on Everett proposed by Price (2010).
Thus, in all these cases (with the exception of Bohm of course), we have a view
in which the probabilities of quantum mechanics are coherently understood as both
purely subjective and ontic, which is what we set out to establish in this section.

3.5 Epistemic Objective Probabilities

We now need to discuss the case of epistemic objective probabilities. We shall do

so in the context of Lewis’s treatment of objective probabilities, which however
immediately raises an obstacle of principle. Recall that in Sect. 3.3 we identified

21 The quantum state is thus taken to be a matter of fact, but presumably not a ‘particular
matter of fact’ in the sense of Lewis, because of the holistic aspects of the quantum state. (Cf.
Footnote 12 above.) Note also that historically the notion of quantum state preceded its ‘statistical
interpretation’, famously introduced by Max Born (for discussion see Bacciagaluppi 2008).
22 For other, even more forceful arguments to this conclusion, see Brown (2019) and Myrvold

(2020).
23 Frigg and Hoefer (2007) have argued that the propensity concept is inadequate to the setting of

GRW. Their own preference is to use ‘Humean objective chances’ instead (cf. Frigg and Hoefer
2010). Note that if the alternative subjectivist strategy includes taking the quantum state as a non-
probabilistic physical object, there are severe constraints on doing this for the case of relativistic
spontaneous collapse theories (see also below, Sects. 3.6 and 3.9).
66 G. Bacciagaluppi

epistemic probabilities as situated in the gap between what is knowable in principle

and what is actually known, and that we settled (at least provisionally) on taking
what is knowable in principle at time t to be the history of the world Ht up to that
time. Now take our credence at t in some proposition A (about some events at time
later than t). As mentioned, crt (A) = cr(A|Et & T ), where Et are the facts at times
up to t that we actually know at that time, and T are our theoretical assumptions. As
in the case of (3.8) above, crt (A) has thus the general form

cr(A|Et & T ) = cr(A|Hti & T )cr(Hti |Et & T ). (3.9)
i

Here, unless the theoretical assumptions T describe the situation as deterministic,

the probabilities cr(A|Hti & T ) will be non-trivial ontic probabilities. (In fact, it
is only now that this statement is unproblematic, after we have argued in the
last section that ontic probabilities can be equally subjective or objective.24 ) The
epistemic component of our credence is given by the various cr(Hti |Et & T ). But if
these probabilities are non-trivial, then according to Lewis they cannot be objective
(i.e. Lewisian chances). Indeed, cr(Hti |Et & T ) is just our credence crt (Hti ) at t
in the proposition Hti . But that is a proposition about past history, and according
to Lewis chances of past events are always 0 or 1. Thus, non-trivial epistemic
probabilities cr(Hti |Et & T ) cannot be objective.
We see that, given the PP, epistemic objective probabilities would seem to be
an oxymoron: if we are ignorant of certain matters of fact, and if propositions about
these facts are admissible in the sense of Lewis, then our non-trivial credences about
these facts are not objective, because chances of past events are 0 or 1. It appears that
Lewis is ruling out by definition that epistemic probabilities could ever be objective.
As seemed to be the case with Popper, also for Lewis epistemic probabilities must
be subjective.
For the sake of arguing that epistemic objective probabilities do make sense after
all, we shall of course grant the existence of objective probabilities in the first place
(pace de Finetti). Thus, assume that e.g. a quantum measurement of electron spin
displays objective probabilities, say 50-50 chances of ‘up’ or ‘down’. We shall take
these as objective, in the Lewisian sense that if I ask you to bet on the outcome of
the next measurement, then conditionally on the chances it is rationally compelling
for you to bet at equal odds.
Now consider the case in which I perform such a measurement but do not show
you the outcome yet, and I ask you to bet. I make the following two claims:
• Your (subjective) probabilities about the outcome are epistemic: there is an
outcome out there which is in principle knowable, and you employ probabilities
only because you do not in fact know it.

24 Indeed, we defined these probabilities as ontic, so unless we accept subjective ontic probabilities,

we are committed to regarding these probabilities either as objective or (more generally!) as

objectively wrong.
3 Unscrambling Subjective and Epistemic Probabilities 67

• You are still going to bet at equal odds, and in fact this is just as rationally
compelling as betting at those odds before the measurement.
The first claim is uncontroversial. The second one – your epistemic probabilities
after the measurement are rationally compelling, granted the probabilities before
the measurement are – I take to be incontrovertible. From a Bayesian perspective,
for instance, you have no new evidence. So it would be irrational to change your
subjective probabilities after I have performed the experiment. But what is more,
I shall now argue that your non-trivial probabilities are objective for the same
reason as in Lewis’s discussion. That is, the rational justification for your probability
assignment is still completely in the spirit of the PP, even though it does not respect
its letter.
Indeed, recall what the PP (3.3) says, as an implicit definition of objective
probabilities: iff there is a proposition X such that, for all choices of reasonable
initial credence function and admissible propositions,
  

cr A  X & Et & T = x, (3.10)

then A has an objective probability and its value is x. Crucially, Lewis sees chances
as functions of time. Thus he requires that our credence function at t should not
incorporate information that is inadmissible at the time t. This corresponds to a
betting situation in which you are allowed to know any matters of fact up to time t,
and the chances fix the objectively fair odds for betting in this situation.
But in the case we are considering now, the betting situation is different. Indeed,
you are not allowed to know what has happened after I have performed the
measurement (even if the measurement was at a time t  < t). If in this betting
situation you knew the outcome of the measurement, that now would be cheating.
In this situation, the proposition that the chance at t is equal to x (in particular 0
or 1) is itself inadmissible information, and it is so for the same reason that Lewis
holds a proposition about the future chance of an event to be inadmissible.
Indeed, in his version of the PP, Lewis is considering a betting context restricted
to what we can in principle know at t (i.e. the past history Ht ), and he asks for the
best we can do in setting our rational credence in a future event A. The answer is
that the best we can do is to set our credence to the chance of A at that time t.
Information about how the chances evolve after t is inadmissible because it illicitly
imports knowledge exceeding what we are allowed to know at t. If we consider
restricting our betting context further so that what we are allowed to know is what
we could in principle have known at an earlier time t  , and now ask for the best
we can do in setting our rational credence in an event at t, then information about
the chance at t of that event has become inadmissible, again precisely because it
exceeds the assumed restriction on what we are allowed to know. The best we can
do is to set our credence to the chance of A at the earlier time t  .
The point is perhaps even clearer if we take Lewis’s reformulated version (3.5)
of the PP:

ch(A|Ht ) = cr(A|Ht & TC ). (3.11)

68 G. Bacciagaluppi

This tells us that if at t we include all admissible information, namely Ht and TC ,

this information already implies what the chances are at t, so that the best we can
do in terms of setting our credence at t, is to take any reasonable initial credence
function and conditionalise on Ht and TC .
But now if we assume that our epistemic context Et is restricted to Ht  , clearly the
best we can do is to take the information Ht  we actually have and feed it into the
complete theory of chance TC , yielding as credence cr(A|Et & TC ) = ch(A|Ht  ).
This credence will now be rationally compelling, even if it is not equal to ch(A|Ht ).
The intuition is that in order to judge whether a subjective probability conditional
on the epistemic context Et = Ht  is in fact objective, we need to compare it not
to the objective probability conditional on the history Ht at the time of evaluation
of our credence, but to the objective probability conditional on the history Ht  that
matches our epistemic context. And this now indeed provides the room for non-
trivial objective probabilities even for events that could in principle be known.

3.6 Redrawing the Subjective-Objective Distinction

We have argued above that one can make good sense both of subjective (indeed,
purely subjective) ontic probabilities and of epistemic objective probabilities, thus
arguing in favour of two unusual conceptual possibilities. But in other respects
we had to be quite conservative. Indeed, the arguments above would have been
weakened had we not tried to keep as much as possible to the accepted use of
the terms ‘epistemic’, ‘ontic’, ‘subjective’ and ‘objective’. Once we have seen
the usefulness of admitting subjective ontic probabilities and epistemic objective
probabilities, however, we might be interested in trying to redraw the subjective-
objective distinction and the epistemic-ontic distinction in ways that are both more
systematic and cleaner in the light of our new insights.
Let us start with the subjective-objective distinction. What we have just argued
for in Sect. 3.5 is that there are more cases of rationally compelling credences than
Lewis in fact allows for.25
Lewis’s intuition is that chances are functions of time, and in the formulation
of the PP this effectively forces him to privilege the maximal epistemic context Ht
comprising all that is knowable in principle at t, or equivalently to quantify over all
epistemic contexts Et comprising what may be actually known at t. Thus, according
to Lewis, we need to ask whether propositions exist that would fix our credences
across all epistemic contexts admissible at time t.
But we have just argued that we can also ask whether propositions exist that
would fix our credences when we are in a particular one of these epistemic contexts.
Specifically, we considered the situation in which at t we only know all matters of

25 Many thanks to Ronnie Hermens for noting that this point was not spelled out clearly enough in
a previous draft.
3 Unscrambling Subjective and Epistemic Probabilities 69

fact up to t  < t. In that case, if we assume Lewisian chances exist, we know in fact
that there are such propositions, namely the chances at t  .
Now we can ask that same question in general: do propositions exist that would
fix our credences when we are in an arbitrary epistemic context E (e. g. any ever so
partial specification of the history up to a time t)? If so, we shall talk of the chances
not at the time t, but the chances given the context E. Lewisian chances at t are then
just the special case when E has the form Ht .
Writing it down in formulas, we have a natural generalisation of the PP and of
objective probabilities: given a context E, iff there is a proposition X such that,
for all choices of reasonable initial credence function and any further theoretical
assumptions T (compatible with X),
  

cr A  X & E & T = x, (3.12)

then A has an objective probability given the context E and its value is x. The
weakest such proposition will be denoted chE (A) = x.
Now, similarly as with Lewis’s own reformulation (3.5) of the PP, note that a
complete theory of chance will now include also history-to-chance conditionals for
any antecedent E for which chances given E exist. Therefore, not only will the
following be an instance of the generalised PP (3.12),
  

cr A  (chE (A) = x) & E & TC = x, (3.13)

but we can further rewrite it as

chE (A) = cr(A|E & TC ). (3.14)

Thus, also our generalised chances will have the kind of properties that Lewis
derives for his own chances from his reformulation (3.5) of the original PP (in
particular, they will obey the probability calculus).
Further, (3.14) makes vivid the intuition that if we have a complete and correct
theory of (generalised) chance, the best we can rationally do if we are in the
epistemic context E, is just to plug the information we actually have into the
complete and correct theory of chance, and let that guide our expectations.
Finally, and very significantly, we see from (3.14) that chances in the more
general sense are not functions of propositions and times, but functions of pairs
of propositions, and they have the form of conditional probabilities. That is,
in this picture chances are fundamentally conditional. I believe this has many
advantages.26

26 Note that also epistemic and ontic probabilities as discussed here are naturally seen as conditional

(on epistemic and ontic contexts, respectively). That all probabilities should be fundamentally
conditional has been championed by a number of authors (see e.g. Hájek 2011 and references
therein). The idea in fact fits nicely with Lewis’s own emphasis that already the notion of possibility
70 G. Bacciagaluppi

From the point of view of the philosophy of physics, such a liberalisation of the
PP means that all kinds of other probabilities used in physics can now be included
in what could potentially be objective chances.
For instance, it allows one to consider the possibility of high-level probabilities
being objective. By high-level probabilities I mean probabilites featuring in theories
such as Boltzmannian statistical mechanics, which is a theory about macrostates,
identified as coarse-grained descriptions of physical systems. Our generalised PP
allows one among other things to apply a Lewisian-style analysis to probabilities
conditional on such a macrostate of a physical system, thus introducing the
possibility of objective chances even in case we are ignorant of the microstate of
a system such as a gas (pace Popper).27
Or take again the EPR setup: Bob performs a measurement, usually taken
to collapse the state also of Alice’s photon, so that the probabilities for her
own measurement results are now cos2 (α/2) and sin2 (α/2). Let us grant that
these probabilities are objective. In Sect. 3.3 we considered Alice’s probability
assignments in the epistemic context in which she does not know yet the result of the
measurement on Bob’s side, and treated them as partially epistemic. Now, however,
we can argue that in Alice’s epistemic context what is rationally compelling for her
is to keep her probabilities unchanged, because we compare them with the quantum
probabilities conditional on events only on her side.
Indeed, the situation is very much analogous to the simple example in which
I perform a measurement but do not tell you the result. If anything, in the EPR
example the case is even more compelling, because there is no way even in principle
that Alice could know Bob’s result if the two measurements are performed at
spacelike separation. Thus, if we take probabilities to be encoded or determined by
the quantum state, we have now made space for the idea that collapse does not act
instantaneously at a distance, but along the future light cone (of Bob’s measurement
in this case).
Quite generally, it is clear that Lewis’s treatment of chances taken literally is
incompatible with relativity, for Lewis’s chances as functions of time presuppose a
notion of absolute simultaneity. In fact, the idea that chance could be relativised to
space-like hypersurfaces or more generally to circumstances which are independent
of absolute time, e.g. properties of space-time regions, was proposed in Berkovitz
(1998), who discusses this proposal with respect to causality and non-locality in the
quantum realm.

is indeed to be naturally understood as a notion of compossibility, of what is possible keeping fixed

a certain context. (Recall his lovely example in Lewis (1976) of whether he can speak Finnish as
opposed to an ape, or cannot as opposed to a native speaker.)
27 And pace Lewis, who explicitly ruled out the possibility of such deterministic chances (Lewis

(1986), p. 118). Note that high-level deterministic chances relate to all kinds of interesting
questions (e.g. emergence, the trade-off between predictive accuracy and explanatory power, the
demarcation between laws and initial conditions, etc.). For recent literature on the debate about
deterministic chances, see e.g. Glynn (2009), Lyon (2011), Emery (2013), and references therein.
3 Unscrambling Subjective and Epistemic Probabilities 71

A relativistic generalisation of Lewisian chances will automatically require more

liberal contexts E – as we have proposed here – because there is no longer an
absolute notion of simultaneity, or of ‘past’, or of ‘becoming’, determining which
events we need to conditionalise upon to temporally evolve our chances. As is well
known, in special relativity these notions can be modified in two ways: one either
takes simultaneity to be relative to inertial frames or more generally to spacelike
foliations, or one localises the notion of absolute past to the past light cone of a
spacetime event (or spacelike segment or bounded spacetime region). Both indeed
require generalising chances from just defining them relative to ‘past histories’ Ht
to contexts that are the past of arbitrary spacelike hyperplanes or hypersurfaces, or,
respectively, to contexts that are the past light cones of arbitrary events (or spacelike
segments or bounded regions).
Conditionalising on the past of arbitrary hypersurfaces is what Myrvold (2000)
has proposed as a way of conceptualising relativistic collapse in the case of quantum
mechanical probabilities.28 Conditionalising on past light cones instead is what
Ismael (2011) has proposed as a way of conceptualising chances in the case of
relativity. Applied to quantum mechanics, this corresponds to the EPR example as
we have just revisited it, with a conception of collapse along the future light cone.29
For yet another example, in recent work Adlam (2018a,b) has argued that
‘temporal locality’ and ‘objective chances’ risk becoming two mutually supporting
dogmas of modern physics. Maybe slightly oversimplifying, the point is that the
idea of objective chances defined at a time t and that of a physical state at time t
being all that is needed to predict the physical state at times t + ε in fact seem to
naturally relate to and support each other. Adlam argues that this creates an obstacle
to the development of e. g. non-Markovian, retrocausal, and ‘all-at-once’ theories.30
But in fact, it is only the standard Lewisian form of objective chances at time t
that may be inimical to such developments, and the present generalisation of the PP
removes this particular obstacle.
For instance, suppose we have some physical theory involving non-Markovian
probabilities. The Lewisian definition of chances actually does allow for non-
Markovian chances, in the sense that cht (A) = ch(A|Ht ) now in general is a
proposition that depends on the whole of the history Ht . However, if at t we only

28 I actually prefer the 2000 archived version to the published paper (Myrvold 2002). More recently,

Myrvold (2017a) has shown that any such theories of foliation-dependent collapse must lead to
infinite energy production. I believe, however, that Myrvold’s idea that probabilities in special
relativity (and in quantum field theory) can be defined conditional on events to the past of an
arbitrary hypersurface can still be implemented (see Sect. 3.9 below).
29 I am indebted to both of these proposals and their authors in more ways than I can say. See also

my Bacciagaluppi (2010a).
30 Retrocausal theories (in particular in the context of quantum mechanics) have been vigorously

championed by Huw Price. All-at-once theories are ones in which probabilities for events are
specified given general boundary conditions, rather than just initial (or final) conditions, and have
been championed in particular by Ken Wharton. See e.g. Price (1997), Wharton (2014) and the
other references in Adlam (2018a). For further discussion see also Rijken (2018) (whom I thank
for discussion of Adlam’s work).
72 G. Bacciagaluppi

have information about some previous times t1 , . . . , tn , the original PP has nothing
to recommend us except to try and figure out the missing information, so as to be
able to determine the chances ch(A|Ht ) at t. Our physical theory, instead, does
contain also probabilities of the form p(A|Et1 & . . . &Etn ), where Eti describes
the state of our physical system at time ti . In order to be able to say that also
these probabilities featuring in our non-Markovian theory are objective, we have
to generalise chances to allow for contexts of the ‘gappy’ form E = Et1 & . . . &Etn
in (3.12)–(3.14).
Note that p(A|Et1 & . . . &Etn ) is just a coarse-graining of probabilities of the
j j
form p(A|Ht ), where Ht is a possible history of the world up to t compatible
with Et1 & . . . &Etn , i.e. a coarse-graining of probabilities which in turn do allow a
reading as Lewisian chances. In this sense, our generalisation (3.12) appears to be
required by sheer consistency, if we want objective probabilities to be closed under
coarse-grainings.31
Finally, even leaving aside motivations from philosophy of physics, what may
perhaps be decisive from the point of view of the philosopny of probability is that
such a generalisation of the PP completely eliminates the need to restrict epistemic
contexts to ‘admissible propositions’. The reason for worrying about admissibility
was that chances were tacitly or explicitly thought to always refer to the context
Ht , and the actual epistemic context had to be restricted to admissible propositions
in order not to bring in any information not contained in Ht . By requiring that
the (ontic) context of the chances exactly match the (epistemic) context of the
credences, no such possibility of mismatch can arise, and no additional conditions
of admissibility are needed.
I suspect Lewis himself might have welcomed such a simplification in the
formulation of the PP, especially given his own qualified formulations when treating
of admissible propositions in connection with the possibility of non-standard
epistemic contexts in the presence of time travel, clairvoyance, and other unusual
cases (as will in fact be the retrocausal and all-at-once theories considered by
Adlam).

3.7 Redrawing the Epistemic-Ontic Distinction

Generalising chances to arbitrary contexts E in our discussion of subjective and

objective probabilities suggests we might want to do something similar in our
discussion of epistemic and ontic probabilities. While sufficient for the purpose of
our examples so far, taking epistemic probabilities to be probabilities of past events
(or any events determined by such events) is now too restrictive. As we did in the

31 Ofcourse also in the case of E = Ht  , the chance given E is the coarse-graining over all
possible intervening histories between t  and t. But this particular coarse-graining does yield again
a Lewisian chance, namely the chance at t  .
3 Unscrambling Subjective and Epistemic Probabilities 73

case of chances, we should not only want to be able to generalise to a relativistic

setting, but also to be able to get rid of the restriction to past events altogether,
allowing in fact for all kinds of non-standard epistemic contexts (clairvoyance,
backwards-directed agents with memories of the future, omniscience, and what not),
and allowing conversely for taking an ontic perspective also on probabilities that are
about past events (e.g. in the case of coarse-graining).
This, however, also seems to raise a problem, namely that the distinction between
epistemic and ontic probabilities becomes a purely pragmatic one at best. If we can
extend or restrict at will which propositions we assume to be knowable in principle,
then any probability can be alternatively seen as epistemic or as ontic, and there is
no substantive difference between the two.
To a large extent, we can bite the bullet, because the distinction between
epistemic and ontic probabilities clearly does have such a pragmatic aspect. When
in Sect. 3.5 we considered the example of the spin measurement with unknown
outcome, in the given epistemic context those probabilities are of course epistemic.
But if we are asked how we choose our probability assignments in that situation, we
will say that we are simply applying the quantum probabilities that obtained at the
time of the measurement, i. e. we are mentally switching to a context in which the
indeterministic outcomes are still before us, and in which the probabilities predicted
by quantum mechanics (which there we assumed are objective) are thus ontic, so
that our theoretical assumptions about that ontic context can inform our probability
assignments in our current epistemic context.
In so doing, however, we also seem to be performing a second switch, one that
was in fact crucial to our arguing in Sect. 3.4 that ontic probabilities could also be
subjective (e.g. when adopting a subjectivist position about probabilities of atomic
decay). Indeed, while epistemic probabilities are about unknown matters of fact in
the actual world, ontic probabilities as we used them in Sect. 3.4 are theoretical
probabilities about alternative possibilities in our models.32
We mentioned already in Footnote 12 above a distinction between material
propositions (propositions about matters of fact in the actual world) and theoretical
propositions (propositions within our theoretical models about the world).33 We
should thus distinguish accordingly between different dimensions of possibility:
material possibilities if we are wondering about what is the case in the material
world, and theoretical possibilities if we are wondering about what could be or might
have been the case.
In this sense, epistemic probabilities are clearly probabilities about material
possibilities: this is the dimension of possibility in which beliefs live, and epistemic
probabilities are measures over unknown matters of fact in the actual world

32 And, indeed, already in Sect. 3.3 we suggested that ontic probabilities tend to be associated with
indeterministic contexts, which we defined in terms of multiple possible futures in our theoretical
models compatible with the present (or present and past) state of the world.
33 As mentioned above, one could of course distinguish further between the actual world and the

material world (if one is a many-world theorist) or take some material propositions to be theoretical
(if one is an empiricist), but for simplicity I neglect these possibilities.
74 G. Bacciagaluppi

(conditional on keeping the epistemic context fixed). But now we recognise an

ambiguity in the standard use of ‘ontic probabilities’ as construed so far: we
have been sliding between on the one hand taking also ontic probabilities to be
material probabilities, namely measures over unknowable facts in the actual world
(conditional on keeping the ontic context fixed, i.e. what is knowable in principle),
and on the other hand taking ontic probabilities to be theoretical probabilities,
namely measures over alternative possible worlds (now thinking of the fixed ontic
context as what we are conditioning on in the model). In the radioactive decay
example this means sliding between taking ontic probabilities as about what the
actual facts in the future are, and taking them as measures over alternative possible
worlds. Our examples of subjective ontic probabilities can be read in either way,
and remain subjective under either reading, whether we identify ontic probabilities
directly with the theoretical probabilities of our models (which in Sect. 3.4 we took
to be purely subjective), or whether we take the ontic probabilities to be material,
and use our subjective theoretical probabilities in fixing them.
It is not essential to choose between the two readings of ontic probabilities, as
long as one can distinguish if required. Insofar as ontic probabilities feature in our
expectations, one might prefer to think of them as material probabilities (since
our expectations are about the actual world). On the other hand, our probabilistic
statements about, say, future quantum events will tend to have modal force (in the
sense of supporting causal and/or law-like statements), indicating that more often
than not we intend such probabilities to be theoretical.34
In this context, while in some of my examples I use the terminology of possible
worlds, or refer to ontic probabilities as law-like, the important thing is indeed that
our theoretical models and the probabilities that feature in them have modal force,
however understood. Our theoretical models may feature probabilities in terms of
probabilistic laws, or in terms of probabilistic causes; if you object to possible
worlds as a formalisation of modality (even though I take them in a very thin sense)
or to thinking of probabilities as law-like, any alternative account of modal force
will do.35
Another point to note is that, while we have said that insofar as epistemic prob-
abilities are material probabilities they do not have modal force, not all theoretical
probabilities need in fact have modal force. Some probabilities in our models may be
understood as contingent elements of a model. Such is typically the case for initial
distributions, or more generally single-time distributions, even in theoretical models
as, say, classical statistical mechanics. And, indeed, when discussing Popper’s

34 Materialprobabilities by definition support only indicative conditionals: ‘If Beethoven was not
born on 16th December, he was born on the 15th’ expresses our next best guess. Instead, when
talking about atomic decay we more commonly use subjunctive or counterfactual conditionals,
such as ‘Should this atom not decay in the next hour, it would have probability 0.5 of decaying
during the following hour’.
35 My thanks to Niels van Miltenburg and to Albert Visser for pressing me on this point.
3 Unscrambling Subjective and Epistemic Probabilities 75

worries in Sect. 3.1 we claimed as self-evident that the probability distributions in

classical statistical mechanics are epistemic. But again we need to distinguish: as
long as we are talking of our credences about what is in fact the case inside a bottle,
these are indeed material and epistemic probabilities; as long, instead, as these are
theoretical probability distributions over different alternative theoretical possibilities
in our model, they do not range over epistemic possibilities. Rather, they appear to
represent the contingent aspects of our theoretical model.36 Thus we see that there is
a certain ambiguity between material and theoretical probabilities also when we are
talking about epistemic probabilities. Again, this ambiguity need not be worrying,
as long as one is clearly aware of the material-theoretical distinction.37
Indeed, to revisit one more time our standard example of mixed probabilities
(unknown composition of a radioactive probe, known decay laws for the various
isotopes), we can now give multiple readings of these probabilities. We can see them
as expressing our expectations about a material proposition in our future, which we
evaluate by mixing our epistemic probabilities for the composition of the probe and
our ontic probabilities (in the material sense) about what will happen in the various
cases. In this case, we fix the values of the latter probabilities using the law-like
probabilities of our theoretical models. Or we see them throughout as probabilities
in a theoretical model, but with a contingent distribution over the isotopes in the
radioactive probe. In that case, we can fix that distribution using our epistemic
probabilities for the composition of the probe. Or yet again, we can think of such
mixed probabilities as a sui generis mix of material and theoretical probabilities,
which we however commonly resort to in practice.
How to use theoretical models (i.e. which probabilities to use) will depend
pragmatically on our epistemic situation. Any given context can be seen as the
epistemic context of some putative agent wondering about the actual world, or as an
ontic context from which to judge the accessibility of other possible worlds. But in
any actual epistemic context in which we wonder about what we do not know in the
material world, we will happily and often resort to theoretical models to guide our
expectations.

36 We shall refine this point in our discussion of time-symmetric theoretical models in Sect. 3.8.
37 It
may, however, contribute to muddle the waters in the Popper and Jaynes examples, or indeed
in modern discussions about ‘typicality’ in statistical mechanics: are we talking about the actual
world (of which there is only one), or about possible worlds (of which there are many, at least in
our models)? It is also not always clear which aspects of a model are indeed law-like and which
are contingent. We shall see a class of examples in Sect. 3.8, but a much-discussed one would be
whether or not the ‘Past Hypothesis’ can be thought of as a law – which it is in the neo-Humean
‘best systems’ approach (cf. Loewer 2007); thanks to Sean Gryb for raising this issue. For further
discussion of typicality and/or the past hypothesis, see e.g. Goldstein (2001), Frigg and Werndl
(2012), Wallace (2011), Pitowsky (2012), and Lazarovici and Reichert (2015).
76 G. Bacciagaluppi

3.8 Remarks on Time Symmetry

Many more things could be said about applying the newly-drawn distinctions in
practice. I shall limit myself in this section to a few remarks on the issue of time
symmetry, and in the next one to some further remarks on the quantum state.
I mentioned in Sect. 3.3 that for the purpose of distinguishing epistemic and ontic
probabilities I did not want to commit myself to a metaphysically thick notion of
indeterminism, with a ‘fixed’ past and an ‘open’ future. Now we can see that such a
notion is in fact incompatible with the way I have suggested drawing the epistemic-
ontic distinction.
Indeed, as drawn in Sect. 3.7, the epistemic-ontic distinction allows for ontic
probabilities also in the case we theoretically assume backwards indeterminism
(where models that coincide at a time t or for times s ≥ t need not coincide at times
earlier than t). We can take the entire history to the future of t as an ontic context
(call it H̃t ), and thus probabilities p(A|H̃t ) about events earlier than t as ontic. This
is plainly incompatible with tying ontic probabilities (whether in the material or the
theoretical sense) to the idea of a fixed past and open future.
There are other reasons for rejecting the idea of fixed past and open future.38
But my suggestion above seems to fly in the face of a number of arguments for a
different metaphysical status of past and future themselves based on probabilistic
considerations. It is these arguments I wish to address here, specifically by briefly
recalling and reframing what I have already written in my Bacciagaluppi (2010b)
against claims that past-to-future conditional probabilities will have a different
status (namely ontic) from that of future-to-past conditional probabilities (allegedly
epistemic). I refer to that paper for details and further references.
Of course we are all familiar with the notion that probabilistic systems can
display strongly time-asymmetric behaviour. To fix the ideas, take atomic decay,
modelled simplistically as a two-level classical Markov process (we shall discuss
quantum mechanics proper in the next section), with an excited state and a de-
excited state, a (comparatively large) probability α for decay from the excited state
to the de-excited state in unit time, and a (comparatively small) probability ε for
spontaneous re-excitation in unit time. We shall take the process to be homogeneous,
i.e. the transition probabilities to be time-translation invariant.
Any initial probability distribution over the two states will converge in time to
an equilibrium distribution in which most of the atoms are de-excited (a fraction
α ε
α+ε ), and only a few are excited (a fraction α+ε ), and the transition rate from the

38 In particular, I agree with Saunders (2002) that standard moves to save ‘relativistic becoming’
in fact undermine such metaphysically thick notions. Rather, I believe that the correct way of
understanding the ‘openness’ of the future, the ‘flow’ of time and similar intuitions is via the logic
of temporal perspectives as developed by Ismael (2017), which straightforwardly generalises to
relativistic contexts, since it exactly captures the perspective of an IGUS (‘Information Gathering
and Utilising System’) along a time-like worldline. For introducing me to the pleasures of time
symmetry I would like to thank in particular Huw Price.
3 Unscrambling Subjective and Epistemic Probabilities 77

excited to the de-excited state will asymptotically match the transition rate from the
de-excited to the excited state (so-called ‘detailed balance’ condition).
This convergence to equilibrium is familiar time-directed behaviour. Of course,
given the joint distributions for two times s > t, transition probabilities can be
defined both from t to s and from s to t (by conditionalising on the state at t and
at s, respectively), but the example can be used in a number of ways to argue that
these transition probabilities forwards and backwards in time must have a different
status. In the language of this paper:
(A) Were backwards probabilities also ontic and equal to the forwards probabilities,
then the system would converge to equilibrium also towards the past, and the
system would have to be stationary (Sober 1993).
(B) While by assumption the forwards transition probabilities are time-translation
invariant, the backwards transition probabilities in general are not, suggesting
that they have a different status (Arntzenius 1995).
(C) If in addition to the forwards transition probabilities also the backwards
transition probabilities were law-like, the initial distribution would also inherit
this law-like status, while it is clearly contingent – indeed can be freely chosen
(Watanabe 1965).
The problem with all these arguments is that they take the observed backwards
frequencies as estimates for the putative ontic backwards probabilities. But fre-
quencies can be used as estimates for probabilities only when we have reasons to
think our samples are unbiased.39 For instance, we would never use post-selected
ensembles to estimate forwards transition probabilities. However, we routinely use
pre-selected ensembles, especially when we set up experiments ourselves and in fact
freely choose the initial conditions. Thus, while we normally use such experiments
to estimate forwards transition probabilities, we plainly cannot rely on them to
estimate the backwards ones.
It turns out that this time-asymmetric behaviour (and a large class of more general
examples) can indeed be modelled using stationary stochastic processes in which
the transition probabilities are time-symmetric, and in which the time asymmetry is
introduced by the choice of special initial (rather than final) frequencies. Thus, in
modelling such time-asymmetric behaviour there is no need to use time-asymmetric
ontic probabilities or to deny ontic status altogether to the backwards transition
probabilities of the process. The apparent asymmetry comes in because we generally
have asymmetric epistemic access to a system: we tend to know aspects of its past
history, and thus are able to pre-select, while its future history will generally be
unknowable.40

39 Thatthis is not always licensed is clear also from discussion of causal loops (Berkovitz 1998,
2001).
40 Of course, this strategy is analogous to that of modelling time-directed behaviour in the

deterministic case using time-symmetric laws and special initial conditions. The same point was
made independently and at the same time by (Uffink 2007, Section 7), specifically in the context
of probabilistic modelling of classical statistical mechanics. A further explicit example of time-
78 G. Bacciagaluppi

If one assumes that the transition probabilities are time-symmetric (or time-
symmetric except for some overall drift41 ), this places severe constraints on the
overall (multi-time) distributions that define the process. Not only is the process
stationary if it is homogeneous, but the process may well be uniquely determined by
its forwards and backwards transition probabilities.42 Thus, in particular, the entire
process inherits the ontic status of the transiton probabilities.
This will seem puzzling, if we are used to thinking of transition probabilities
as ontic, and of initial distributions as epistemic, so that the probabilities of the
process should have a mixed character (as in the examples we discussed beginning
in Sect. 3.3). But as we saw in Sect. 3.7, we standardly align the distinction between
epistemic and ontic probabilities with the distinction between contingent and law-
like aspects of our theoretical probabilistic models. The theoretical probabilities
of the process may very well be even uniquely determined by the forwards and
backwards transition probabilities, but our theoretical models will still provide scope
for contingent elements through the notion of samples of the process. We then
feed our epistemic probabilities into the theoretical model via constraints on the
frequencies in the samples, which will typically be initial constraints because of our
time-asymmetric access to the empirical world.43

3.9 Further Remarks on the Quantum State

We have just seen that, classically, a stochastic process may be nothing over and
above an encoding of its forwards and backwards transition probabilities. I now

symmetric ontic probabilities is provided by Nelson’s (1966,1985) stochastic mechanics, where the
quantum mechanical wavefunction and the Schrödinger equation are derived from an underlying
time-symmetric diffusion process. See also Bacciagaluppi (2005) and references therein. The
notorious gap in Nelson’s derivation pointed out by Wallstrom (1989) has now been convincingly
filled through the modification of the theory recently proposed by Derakhshani (2017).
41 In the case of Nelson’s mechanics, for instance, one has a systematic drift velocity that equals

the familiar de Broglie-Bohm velocity (which in fact also changes sign under time reversal). If one
subtracts the systematic component, transition probabilities are then exactly time-symmetric.
42 More precisely, it will be uniquely determined within each ‘ergodic class’. (Note that ergodicity

results are far easier to prove for stochastic processes than for deterministic dynamical systems.
See again the references in Bacciagaluppi (2010b) for details.)
43 Without going into details, these distinctions will be relevant to debates about the ‘neo-

Boltzmannian’ notion of typicality (see the references in Footnote 37), and about the notion of
‘sub-quantum disequilibrium’ in theories like pilot-wave theory or Nelsonian mechanics (for the
former see e.g. Valentini 1996, for the latter see Bacciagaluppi 2012).
3 Unscrambling Subjective and Epistemic Probabilities 79

wish to briefly hint at the possibility that the same may be true also in quantum
mechanics.44
In quantum mechanics, the role of determining transition probabilities belongs
to the quantum state (either the fixed Heisenberg-picture state, or the totality of
the time-evolving Schrödinger-picture or interaction-picture states45 ). The quantum
state, however, is traditionally seen as more than simply encoding transition
probabilities, rather as a physical object in itself, which in particular determines
these probabilities. One obvious exception is qBism, where however I have argued
in Sect. 3.4 above that the rejection of the quantum state as a physical object is not
sufficiently motivated.46
In this section I want to take a look at a different approach to quantum theory,
where I believe there are strong constraints to taking the quantum state as an
independently existing object, and reasons to take it indeed as just encoding (ontic,
law-like) transition probabilities, largely independently of the issue of whether these
probabilities should be subjective or objective.47
This approach is the foliation-dependent approach to relativistic collapse. As
is well-known, non-relativistic formulations of collapse will postulate that the
quantum state collapses instantaneously at a distance (whether the collapse is
spontaneous or the traditional collapse upon measurement). So, for instance, in an
EPR scenario the collapse will depend on whether Alice’s or Bob’s measurement
takes place first. However, predictions for the outcomes of the measurements are
independent of their order because the measurements commute: we obtain the
same probabilities for the same pairs of measurement results. This suggests that
one might be able to obtain a relativistically invariant formulation of collapse
after all. One of the main strategies for attempting this takes quantum states to
be defined relative to spacelike hyperplanes or more generally relative to arbitrary
spacelike hypersurfaces, and collapse accordingly to be defined relative to spacelike
foliations. This option has been pioneered for a long time by Fleming (see e.g.

44 This section can be skipped without affecting the rest of the paper. The discussion is substantially

influenced by Myrvold (2000), but should be taken as an elaboration rather than exposition of
his views. See also Bacciagaluppi (2010a). For Myrvold’s own recent views on the subject, see
Myrvold (2017b, 2019).
45 The interaction picture is obtained from the Schrödinger picture by absorbing the free evolution

into the operators. In other words, one encodes the free terms of the evolution in the evolution of
the operators, and encodes the interaction in the evolution of the state.
46 Another obvious exception is Nelson’s stochastic mechanics (mentioned in Footnote 40), but

there it is not the quantum mechanical transition probabilities that are fundamental but the
transition probabilities of the underlying diffusion process. While also a number of Bohmians (cf.
Goldstein and Zanghì 2013) think of the wavefunction not as a material entity but as nomological,
as a codification of the law-like motion of the particles, there is no mathematical derivation in
pilot-wave theory of the wavefunction and Schrödinger equation from more fundamental entities
comparable to that in stochastic mechanics, and the wavefunction can be thought of as derivative
only if one takes a quite radical relationist view, like the one proposed by Esfeld (forthcoming).
47 For a sketch of how I think this relates to the PBR theorem (Pusey et al. 2012), see below.
80 G. Bacciagaluppi

1986,1989,1996), and analysed conceptually in great detail by Myrvold (2000) (see

also Bacciagaluppi 2010a).48
A convenient formal framework for foliation-dependent collapse is provided
by the Tomonaga-Schwinger formalism of quantum field theory, in which the
(interaction-picture) quantum state is explicitly hypersurface-dependent, and unitary
evolution in the presence of interactions is represented as a local evolution of the
quantum state from one hypersurface to the next (and is independent of the choice
of foliation between two hypersurfaces S and S  ). A non-unitary collapse evolution
can be included in the same way, as proposed explicitly e.g. by Nicrosini and
Rimini (2003), provided one can ensure that the evolution is also independent of
the foliation chosen.
The substantial difference between the unitary and non-unitary case is that in
the unitary case, the local state on a segment of a spacelike hypersurface (say, the
reduced state of Alice’s electron) is independent of whether the segment, say, T ∩T 
is considered to be a segment of a hypersurface T or of a different hypersurface T  .
In the non-unitary case, in general the restriction ψ T |T ∩T  on T ∩ T  of a state ψ T
on T is different from the restriction ψ T  |T ∩T  on T ∩ T  of the state ψ T  on T  .
Indeed, in the EPR case ψ T could be the singlet state, so that ψ T |T ∩T  is maximally
mixed, while T  could be to the future of Bob’s measurement, so that ψ T  |T ∩T  is,
say, the state |+y . In particular, whether Alice’s electron is entangled with Bob’s
will depend on the hypersurface one considers (or the frame of reference – Myrvold
(2000) aptly calls this the ‘relativity of entanglement’).
This now raises issues of how to interpret such a collapse theory, specifically
in a way that could give rise to the manifest image of localised objects in space
and time. The standard options for doing so go under the names of ‘wavefunction
ontology’, ‘mass-density ontology’, and ‘flash ontology’, depending on whether one
takes the physical objects of the theory to be the quantum states, local mass densities
defined via the states, or the collapse events themselves. Hypersurface-dependence
of quantum states appears to rule out the wavefunction ontology, precisely because
if quantum states are hypersurface-dependent, they do not have unique local
restrictions, so an ontology based on them fails to underpin any local reality.
Similarly, mass-density ontology is ruled out, because hypersurface-dependent
states do not define uniquely any local mass densities. The only option that could
give rise to a localised manifest image of reality appears to be the flash ontology.
This means that hypersurface-dependent states indeed merely encode probabilities
for physical events to the future of a spacelike hypersurface, conditional on physical
events to the past of that hypersurface. If so, as Myrvold (2000, Footnote 7) remarks,
it is unsatisfactory that we should not yet have a formulation of the theory directly
in terms of transition probabilities, but require an additional theoretical object, the

48 The other main strategy is to consider collapse to be occurring along light cones (either along the

future light cone or along the past light cone, as in the proposal by Hellwig and Kraus (1970)), and
it arguably allows one to retain the picture of quantum states as physical objects (Bacciagaluppi
2010a). No such theory has been developed in detail, however.
3 Unscrambling Subjective and Epistemic Probabilities 81

quantum state, to determine these probabilities. (As we saw in Sect. 3.6, also Adlam
(2018a,b) is critical of thinking of a separate physical object as determining the
quantum probabilities.)
What I wish to suggest is that, as in the classical case of Sect. 3.8, the use of
a separate quantum state to calculate transition probabilities may be an artefact of
our own limited epistemic perspective, and that one needs no more than forwards
and backwards transition probabilities corresponding to a ‘stationary process’. The
form of this process is easy to guess, reasoning by time symmetry from the fact that
the future boundary condition we standardly use is the maximally mixed state. If
we take that to be the case because we happen to lack information about the future,
and therefore do not post-select, that suggests that if we want to remove the pre-
selection that we normally do, and thus get to the correct theoretical probabilities,
we need to take the maximally mixed state also as the past boundary condition. The
appearance of a contingent quantum state would then follow from the fact that we
have epistemic access to past collapses, and conditionalise upon them when making
predictions, in general mixing our epistemic probabilities about such past collapses
(‘results of measurements’) and ontic transition probabilities provided by the theory.
Of course in general the maximally mixed state is unnormalisable, but condi-
tional probabilities will be well-defined. In such a ‘no-state’ quantum theory not
only are conditional probabilities fundamental, but unconditional probabilities in
general do not even exist.49
No-state foliation-dependent flash-ontology quantum field theory would thus be
a way of directly implementing Myrvold’s idea that conditional probabilities should
be determined by the theory directly, and not by way of the quantum state. Better
predictions will be obtained the more events one conditionalises upon stretching
back into the past. The theory is thus essentially non-Markovian (thus temporally
non-local, in line with Adlam’s views). As mentioned in Footnote 28, Myrvold
(2017a), developing remarks by Pearle (1993), has shown that foliation-dependent
collapse theories always run into a problem of infinite energy production, but I
believe a no-state theory in fact escapes this conclusion.50

49 Rather related ideas are common in the decoherent histories or consistent histories literature.
Griffiths (1984) does not use an initial and a final state, but an initial and a final projection of the
same kind as the other projections in a history, so his consistent histories formalism is in fact a
no-state quantum theory. Hartle (1998) suggests that the fundamental formula for the decoherence
functional is the two-state one, but that we are well-shielded from the future boundary condition so
that two-state quantum theory is predictively equivalent to standard one-state quantum theory (i.e.
we can assume the maximally mixed state as future boundary condition). If in one-state quantum
theory we assume that we are similarly shielded from the initial boundary condition, then by
analogy one-state quantum theory is in fact predictively equivalent to no-state quantum theory
(i.e. we can assume the maximally mixed state also as past boundary condition). Thus, it could just
as well be that no-state quantum theory is fundamental. (Of course we need to include pre-selection
on the known preparation history.)
50 The idea behind Myrvold’s proof is extremely simple. Leaving technicalities aside, because of

the Lorentz-invariance of the vacuum state, infinite energy production would arise if localised
collapse operators did not leave the vacuum invariant on average. Thus, one should require that
82 G. Bacciagaluppi

Such a proposal has in fact been put forward in the non-relativistic case by
Bedingham and Maroney (2017), in a paper in which they discuss in general
the possibility of time-symmetric collapse theories. As with the requirement of
Lorentz invariance, the requirement of time symmetry makes it problematic to
think of the collapsing quantum state as a physical object, because the sequence
of collapsing states defined by the collapse equations applied in the past-to-future
direction is not the time reversal of the sequence of states defined by the collapse
equations applied in the future-to-past direction. That is, the quantum state is time-
direction-dependent. However, under very mild conditions, from the application of
the equations in the two directions of time one does obtain the same probabilities for
the same collapse events. This suggests, again as in the relativistic case, that a time-
symmetric theory of collapse requires a flash ontology. And, as I am suggesting in
the case of hypersurface-dependence, Bedingham and Maroney suggest one might
want to get rid altogether of the quantum state (since the collapse mechanism
suggests it is asymptotically maximally mixed).51
One can find an analogy for the approach discussed in this section by looking at
Heisenberg’s views on ‘quantum mechanics’ (i.e. matrix mechanics) as expressed in
the 1920s in opposition to Schrödinger’s views on wave mechanics.52 The evidence
suggests that Heisenberg did not believe in the existence of the quantum state as a
physical object, or of collapse as a physical process. Rather, the original physical
picture behind matrix mechanics was one in which quantum systems performed

local collapse operators leave the vacuum invariant. But since a local operator commutes with
local operators at spacelike separation, and since (by the Reeh-Schlieder theorem) applications
of the latter can approximate any global state, it follows that in order to avoid infinite energy
production collapse operators must leave every state invariant. Thus there is no collapse.
One way out that is already present in the literature (and which motivates Myrvold’s theorem)
is to tie the collapse to some ‘non-standard’ fields (which commute with themselves at spacelike
and timelike separation), as done by both Bedingham (2011) and Pearle (2015) himself. Another
possibility, favoured by Myrvold (2018a), is to interpret the result as supporting the idea that
collapse is tied to gravity: in the curvature-free case of Minkowski spacetime, the theorem shows
there is no collapse. As mentioned, however, I believe that a no-state proposal will also circumvent
the problem of infinite energy production, because collapse operators will automatically leave the
(Lorentz-invariant) maximally mixed state invariant on average.
51 I should briefly remark on how I think these suggestions relate to the work of Pusey et al. (2012),

who famously draw an epistemic-ontic distinction for the quantum state, and who conclude from
their PBR theorem that the quantum state is ontic. An ontic state for PBR is an independently
existing object that determines the ontic probabilities (in my sense) for results of measurements.
And an epistemic reading of the quantum state for PBR is a reading of it as an epistemic probability
distribution (in my sense) over the ontic states of a system. Thus, the PBR theorem seems to
push for a wavefunction ontology. In a no-state proposal (as here or in Bedingham and Maroney
2017), however, the probabilities for future (or past) collapse events are fully determined by the
set of past (or future) collapse events. The contingent quantum state (in fact whether defined along
hypersurfaces or along lightcones) can thus be identified with such a set of events, which is indeed
ontic, thus removing the apparent contradiction. Whether this is the correct way of thinking about
this issue of course deserves further scrutiny.
52 For more detailed accounts, see Bacciagaluppi and Valentini (2009), Bacciagaluppi (2008),

Bacciagaluppi and Crull (2009), and Bacciagaluppi et al. (2017).

3 Unscrambling Subjective and Epistemic Probabilities 83

quantum jumps between stationary states. This then evolved into a picture of
quantum systems performing stochastic transitions between values of measured
quantities. Transition probabilities for Heisenberg were ‘out there’ in the world (i.e.
presumably objective as well as ontic). If a quantity had been measured, then it
would perform transitions according to the quantum probabilities. The observed
statistics would conform to the usual rules of the probability calculus, in particular
the law of total probability. But if a quantity had not been measured, there would be
nothing to perform the transitions, giving rise to the phenomenon of ‘interference of
probabilities’.
Schrödinger’s quantum states were for Heisenberg nothing but a convenient tool
for calculating (forwards) transition probabilities, and any state that led to the correct
probabilities could be used. (Note that, classically, forwards transition probabilities
are independent of the chosen initial conditions.) Hence in particular, the movability
of the ‘Heisenberg cut’ and Heisenberg’s cavalier attitude to collapse,53 and the
peculiar mix of ‘subjective’ and ‘objective’ elements in talking about quantum
probabilities. And now, finally, despite the shrouding in ‘Copenhagen mist’, we can
perhaps start to make sense of some of Heisenberg’s intuitions as expressed in ‘The
Copenhagen interpretation of quantum theory’ (Heisenberg 1958, Chap. 3).

3.10 Hume and de Finetti

This paper has suggested a framework for thinking of probabilities in a way that
distinguishes between subjective and epistemic probabilities. Two strategies have
been used: liberalising the contexts with respect to which one can define various
kinds of probabilities, and making a liberal use of probabilities over theoretical
possibilities. I wish to conclude by putting the latter into a wider perspective and
very briefly sketching the kind of subjectivist and anti-metaphysical position that I
wish to endorse.54
In forming theoretical models, whether they involve causal, probabilistic or any
other form of reasoning, we aim at describing the behaviour of our target systems in
a way that guides our expectations. To do so, our models need to have modal force,
to describe law-like rather than accidental behaviour. It is up for grabs what this
modal force is. We can have an anti-Humean view of causes, laws or probabilities,
as some kind of necessary connections, but Hume (setting aside the finer scruples
of Hume scholarship) famously denied the existence of such necessary connections.
This traditionally gives rise to the problem of induction. Hume himself proposed

53 For an example of (Born and) Heisenberg using a very unusual choice of ‘collapsed states’ to
calculate quantum probabilities, see (Bacciagaluppi and Valentini 2009, Sect. 6.1.2).
54 To this position, I would wish to add a measure of empiricism (Bacciagaluppi 2019). For a

similar view, see Brown and Ben Porath (2020). On these matters (and several others dealt with in
this paper) I am earnestly indebted to many discussions with Jenann Ismael over the years. Any
confusions are entirely my own, however.
84 G. Bacciagaluppi

a ‘sceptical solution’: we are predisposed to come to consider certain connections

between events as causal through a process of habituation, which then forms the
basis of our very sophisticated reasonings about matters of fact (‘Elasticity, gravity,
cohesion of parts, communication of motion by impulse; these are probably the
ultimate causes and principles which we shall ever discover in nature; and we may
esteem ourselves sufficiently happy, if, by accurate enquiry and reasoning, we can
trace up the particular phenomena to, or near to, these general principles’, Hume
1748, IV.I.26). There is no guarantee that our expectations will come out true, but
these forms of reasoning have served us well.
There is no need, however, for Humeans to restrict themselves to thinking of
theoretical models in terms of Hume’s own analysis of causation. Any theoretical
models are indeed patterns we project onto the world of matters of fact in order to
reason about it. Modern-day pragmatists about causation are essentially following
Hume, even though they substitute manipulation for habituation (Menzies and
Price 1993). And the same is true of subjectivists about probability. Subjective
probabilities are patterns that we form in our head, and impose on matters of fact to
try and order them for our pragmatic purposes.
Self-styled latter-day Humeans often attempt to give objective underpinnings to
the laws and probabilities in our models. The best systems approach takes Humean
laws to be the axioms in the best possible systematisation of the totality of events,
and Lewis’s ‘Subjectivist Guide’ is a blueprint for the objective underpinning of
subjective probabilities.
But the Humean may well have no need of such objective underpinnings. Radical
subjectivists about probability have been saying so for a long time. There is no sense
in which probabilistic judgements are right or wrong, just as there is no necessary
connection between causes and effects. By and large through a process similar to
habituation (Bayesian conditioning!) we form the idea of probabilistic connections,
which then forms the basis of our further reasonings about matters of fact. There is
no guarantee that our expectations will come out true, but these forms of reasoning
have served us well.
For the subjectivist, or pragmatist, such are all our theoretical models of science,
including quantum mechanics. And we judge them in terms of the standard criteria
of science: first and foremost the empirical criterion of past performance, as well as
other criteria such as simplicity, expectation of fruitfulness, and so on. In this sense,
there are no objective chances in the sense of strict Lewisian rationality, no objective
laws nor objective modality in either an anti-Humean or even neo-Humean sense,
but merely in the sense of the pragmatic rationality of science.
Indeed, the best systems analysis provides a plausible if schematic picture of our
inductive practices, but there is no point to applying it at the ‘end of time’, when
we no longer need to project any predicates. Any laws that we formulate along
the way, any notions of modality that we construe are only the result of analysing
the epistemically available data, and as such are only projections of what has been
successful so far. From a Humean point of view, this seems to me as it should be, just
as Hume himself was content with his sceptical solution to the problem of induction.
3 Unscrambling Subjective and Epistemic Probabilities 85

And, as a matter of fact, in the relatively neglected Section VI of the Enquiry,

Hume explicitly discusses probabilities in terms of forming expectations, again
through habit, about what will happen in certain proportions of cases. Thus Hume
was perhaps the first de Finettian. . .

Acknowledgements I am deeply grateful to Meir Hemmo and Orly Shenker for the invitation to
contribute to this wonderful volume, as well as for their patience. I have given thanks along the way,
both for matters of detail and for inspiration and stimuli stretching back many years, but I must add
my gratitude to students at Aberdeen, Utrecht and the 5th Tübingen Summer School in the History
and Philosophy of Science, where I used bits of this material for teaching, and to the audience at
two Philosophy of Science seminars at Utrecht, where I presented earlier versions of this paper,
in particular to Sean Gryb, Niels van Miltenburg, Wim Mol, Albert Visser and Nick Wiggershaus.
Special thanks go to Ronnie Hermens for precious comments on the first written draft, further
thanks to Harvey Brown, Alan Hájek and Jenann Ismael for comments and encouraging remarks
on the final draft, and very special thanks to Jossi Berkovitz for a close reading of the final draft
and many detailed suggestions and comments.

References

Adlam, E. (2018a). Spooky action at a temporal distance. Entropy, 20(1), 41–60.

Adlam, E. (2018b). A tale of two anachronisms. Plenary talk given at Foundations 2018, Utrecht
University, 13 July 2018. https://foundations2018.sites.uu.nl/.
Arntzenius, F. (1995). Indeterminism and the direction of time. Topoi, 14, 67–81.
Bacciagaluppi, G. (2005). A conceptual introduction to Nelson’s mechanics. In R. Buccheri, M.
Saniga, & E. Avshalom (Eds.), Endophysics, time, quantum and the subjective (pp. 367–388).
Singapore: World Scientific. Revised version at http://philsci-archive.pitt.edu/8853/.
Bacciagaluppi, G. (2008). The statistical interpretation according to Born and Heisenberg. In
C. Joas, C. Lehner, & J. Renn (Eds.), HQ-1: Conference on the history of quantum physics
(MPIWG preprint series, Vol. 350, pp. 269–288). Berlin: MPIWG. http://www.mpiwg-berlin.
mpg.de/en/resources/preprints.html.
Bacciagaluppi, G. (2010a). Collapse theories as beable theories. Manuscrito, 33(1), 19–54. http://
philsci-archive.pitt.edu/8876/.
Bacciagaluppi, G. (2010b). Probability and time symmetry in classical Markov processes. In M.
Suárez (Ed.), Probabilities, causes and propensities in physics (Synthese library, Vol. 347,
pp. 41–60). Dordrecht: Springer. http://philsci-archive.pitt.edu/archive/00003534/.
Bacciagaluppi, G. (2012). Non-equilibrium in stochastic mechanics. Journal of Physics: Confer-
ence Series, 361(1), article 012017. http://philsci-archive.pitt.edu/9120/.
Bacciagaluppi, G. (2014). A critic looks at qBism. In M. C. Galavotti, D. Dieks, W. Gonzalez, S.
Hartmann, T. Uebel, & M. Weber (Eds.), New directions in the philosophy of science (pp. 403–
416). Cham: Springer. http://philsci-archive.pitt.edu/9803/.
Bacciagaluppi, G. (2016). Quantum probability – An introduction. In A. Hájek & C. Hitchcock
(Eds.), The Oxford handbook of probability and philosophy (pp. 545–572). Oxford: Oxford
University Press. http://philsci-archive.pitt.edu/10614/.
Bacciagaluppi, G. (2019). Adaptive empiricism. In G. M. D’Ariano, A. Robbiati Bianchi, & S.
Veca (Eds.), Lost in physics and metaphysics – Questioni di realismo scientifico (pp. 99–113).
Milano: Istituto Lombardo Accademia di Scienze e Lettere. https://doi.org/10.4081/incontri.
2019.465.
Bacciagaluppi, G., & Crull, E. (2009). Heisenberg (and Schrödinger, and Pauli) on hidden
variables. Studies in History and Philosophy of Modern Physics, 40, 374–382. http://philsci-
archive.pitt.edu/archive/00004759/.
86 G. Bacciagaluppi

Bacciagaluppi, G., & Valentini, A. (2009). Quantum theory at the crossroads – Reconsidering the
1927 Solvay conference. Cambridge: Cambridge University Press.
Bacciagaluppi, G., Crull, E., & Maroney, O. (2017). Jordan’s derivation of blackbody fluctuations.
Studies in History and Philosophy of Modern Physics, 60, 23–34. http://philsci-archive.pitt.
edu/13021/.
Bedingham, D. (2011). Relativistic state reduction dynamics. Foundations of Physics, 41, 686–704.
Bedingham, D., & Maroney, O. (2017). Time reversal symmetry and collapse models. Foundations
of Physics, 47, 670–696.
Bell, J. S. (1984). Beables for quantum field theory. Preprint CERN-TH-4035. Reprinted in: (1987).
Speakable and unspeakable in quantum mechanics (pp. 173–180). Cambridge: Cambridge
University Press.
Berkovitz, J. (1998). Aspects of quantum non-locality II – Superluminal causation and relativity.
Studies in History and Philosophy of Modern Physics, 29(4), 509–545.
Berkovitz, J. (2001). On chance in causal loops. Mind, 110(437), 1–23.
Berkovitz, J. (2012). The world according to de Finetti – On de Finetti’s theory of probability and
its application to quantum mechanics. In Y. Ben-Menahem & M. Hemmo (Eds.), Probability
in physics (pp. 249–280). Heidelberg: Springer.
Berkovitz, J. (2019). On de Finetti’s instrumentalist philosophy of probability. European Journal
for Philosophy of Science, 9, article 25.
Brown, H. (2019). The reality of the wavefunction – Old arguments and new. In A. Cordero
(Ed.), Philosophers look at quantum mechanics (Synthese library, Vol. 406, pp. 63–86). Cham:
Springer.
Brown, H., & Ben Porath, G. (2020). Everettian probabilities, the Deutsch-Wallace theorem and
the principal principle. This volume, pp. 165–198.
Bub, J. (2007). Quantum probabilities as degrees of belief. Studies in History and Philosophy of
Modern Physics, 38, 232–254.
Bub, J. (2010). Quantum probabilities – An information-theoretic interpretation. In S. Hartmann
& C. Beisbart (Eds.), Probabilities in physics (pp. 231–262). Oxford: Oxford University Press.
Bub, J., & Pitowsky, I. (1985). Critical notice – Sir Karl R. Popper, Postscript to the logic of
scientific discovery. Canadian Journal of Philosophy, 15(3), 539–552.
de Finetti, B. (1970). Theory of probability. New York: Wiley.
Derakhshani, M. (2017). Stochastic mechanics without ad hoc quantization – Theory and
applications to semiclassical gravity. Doctoral dissertation, Utrecht University. https://arxiv.
org/pdf/1804.01394.pdf.
Diaconis, P., Holmes, S., & Montgomery, R. (2007). Dynamical bias in the coin toss. SIAM Review,
49(2), 211–235.
Emery, N. (2013). Chance, possibility, and explanation. The British Journal for the Philosophy of
Science, 66(1), 95–120.
Esfeld, M. (forthcoming). A proposal for a minimalist ontology. Forthcoming in Synthese, https://
doi.org/10.1007/s11229-017-14268.
Fleming, G. (1986). On a Lorentz invariant quantum theory of measurement. In D. M. Greenberger
(Ed.), New techniques and ideas in quantum measurement theory (Annals of the New York
Academy of Sciences, Vol. 480, pp. 574–575). New York: New York Academy of Sciences.
Fleming, G. (1989). Lorentz invariant state reduction, and localization. In A. Fine & M. Forbes
(Eds.), PSA 1988 (Vol. 2, pp. 112–126). East Lansing: Philosophy of Science Association.
Fleming, G. (1996). Just how radical is hyperplane dependence? In R. Clifton (Ed.), Perspectives
on quantum reality – Non-relativistic, relativistic, and field-theoretic (pp. 11–28). Dordrecht:
Kluwer Academic.
Frigg, R., & Hoefer, C. (2007). Probability in GRW theory. Studies in History and Philosophy of
Modern Physics, 38(2), 371–389.
Frigg, R., & Hoefer, C. (2010). Determinism and chance from a Humean perspective. In D. Dieks,
W. Gonzalez, S. Hartmann, M. Weber, F. Stadler, & T. Uebel (Eds.), The present situation in
the philosophy of science (pp. 351–371). Berlin: Springer.
Frigg, R., & Werndl, C. (2012). Demystifying typicality. Philosophy of Science, 79(5), 917–929.
3 Unscrambling Subjective and Epistemic Probabilities 87

Fuchs, C. A. (2002). Quantum mechanics as quantum information (and only a little more). arXiv
preprint www.quant-ph/0205039.
Fuchs, C. A. (2014). Introducing qBism. In M. C. Galavotti, D. Dieks, W. Gonzalez, S. Hartmann,
T. Uebel, & M. Weber (Eds.), New directions in the philosophy of science (pp. 385–402). Cham:
Springer.
Glynn, L. (2009). Deterministic chance. The British Journal for the Philosophy of Science, 61(1),
51–80.
Goldstein, S. (2001). Boltzmann’s approach to statistical mechanics. In J. Bricmont, D. Dürr,
M. C. Galavotti, G.C. Ghirardi, F. Petruccione, & N. Zanghì (Eds.), Chance in physics (Lecture
notes in physics, Vol. 574, pp. 39–54). Berlin: Springer.
Goldstein, S., & Zanghì, N. (2013). Reality and the role of the wave function in quantum theory.
In A. Ney, & D. Z. Albert (Eds.), The wave function – Essays on the metaphysics of quantum
mechanics (pp. 91–109). Oxford: Oxford University Press.
Greaves, H., & Myrvold, W. (2010). Everett and evidence. In S. Saunders, J. Barrett, A. Kent, &
D. Wallace (Eds.), Many worlds? Everett, quantum theory, and reality (pp. 264–304). Oxford:
Oxford University Press.
Griffiths, R. B. (1984). Consistent histories and the interpretation of quantum mechanics. Journal
of Statistical Physics, 36(1–2), 219–272.
Hájek, A. (2011). Conditional probability. In D. M. Gabbay, P. Thagard, J. Woods, P. S.
Bandyopadhyay, & M. R. Forster (Eds.), Philosophy of statistics, Handbook of the philosophy
of science (Vol. 7, pp. 99–135). Dordrecht: North-Holland.
Hartle, J. B. (1998). Quantum pasts and the utility of history. Physica Scripta T, 76, 67–77.
Heisenberg, W. (1958). Physics and philosophy – The revolution in modern science. New York:
Harper.
Hellwig, K.-E., & Kraus, K. (1970). Formal description of measurements in local quantum field
theory. Physical Review D, 1, 566–571.
Howard, D. (2004). Who invented the Copenhagen interpretation? A study in mythology.
Philosophy of Science, 71(5), 669–682.
Hume, D. (1748). An enquiry concerning human understanding. London: A. Millar.
Ismael, J. (2011). A modest proposal about chance. The Journal of Philosophy, 108(8), 416–442.
Ismael, J. (2017). Passage, flow, and the logic of temporal perspectives. In C. Bouton & P. Huneman
(Eds.), Time of nature and the nature of time (Boston studies in the philosophy and history of
science, Vol. 326, pp. 23–38). Cham: Springer.
Jaynes, E. T. (1985). Some random observations. Synthese, 63(1), 115–138.
Lazarovici, D., & Reichert, P. (2015). Typicality, irreversibility and the status of macroscopic laws.
Erkenntnis, 80(4), 689–716.
Lewis, D. (1976). The paradoxes of time travel. American Philosophical Quarterly, 13(2), 145–
152.
Lewis D. (1980). A subjectivist’s guide to objective chance. In R. C. Jeffrey (Ed.), Studies in
inductive logic and probability (Vol. 2, pp. 263–293). Reprinted in: (1986). Philosophical
papers (Vol. 2, pp. 83–113). Oxford: Oxford University Press.
Lewis, D. (1986). Postscripts to ‘A subjectivist’s guide to objective chance’. In Philosophical
papers (Vol. 2, pp. 114–132). Oxford: Oxford University Press.
Loewer, B. (2004). David Lewis’s Humean theory of objective chance. Philosophy of Science, 71,
1115–1125.
Loewer, B. (2007). Counterfactuals and the second law. In H. Price & R. Corry (Eds.), Causation,
physics, and the constitution of reality – Russell’s republic revisited (pp. 293–326). Oxford:
Oxford University Press.
Lyon, A. (2011). Deterministic probability – Neither chance nor credence. Synthese, 182(3), 413–
432.
Menzies, P., & Price, H. (1993). Causation as a secondary quality. British Journal for the
Philosophy of Science, 44, 187–203.
Myrvold, W. (2000). Einstein’s untimely burial. PhilSci preprint http://philsci-archive.pitt.edu/
222/.
88 G. Bacciagaluppi

Myrvold, W. (2002). On peaceful coexistence – Is the collapse postulate incompatible with

relativity? Studies in History and Philosophy of Modern Physics, 33(3), 435–466.
Myrvold, W. (2017a). Relativistic Markovian dynamical collapse theories must employ nonstan-
dard degrees of freedom. Physical Review A, 96(6), 062116.
Myrvold, W. (2017b). Ontology for collapse theories. In Shan Gao (Ed.), Collapse of the wave
function – Models, ontology, origin, and implications (pp. 99–126). Cambridge: Cambridge
University Press.
Myrvold, W. (2018a). Private communication, New Directions in Philosophy of Physics, Viterbo,
June 2018. http://carnap.umd.edu/philphysics/newdirections18.html.
Myrvold, W. (2018b). Philosophical issues in quantum theory. In E. N. Zalta (Ed.), The Stanford
encyclopedia of philosophy (Fall 2018 Edition). https://plato.stanford.edu/archives/fall2018/
entries/qt-issues/.
Myrvold, W. (2019). Ontology for relativistic collapse theories. In O. Lombardi, S. Fortin, C.
López, & F. Holik (Eds.), Quantum worlds (pp. 9–31). Cambridge: Cambridge University
Press. http://philsci-archive.pitt.edu/14718/.
Myrvold, W. (2020). Subjectivists about quantum probabilities should be realists about quantum
states. This volume, pp. 449–465.
Nelson, E. (1966). Derivation of the Schrödinger equation from Newtonian mechanics. Physical
Review, 150, 1079–1085.
Nelson, E. (1985). Quantum fluctuations. Princeton: Princeton University Press.
Nicrosini, O., & Rimini, A. (2003). Relativistic spontaneous localization – A proposal. Founda-
tions of Physics, 33(7), 1061–1084.
Pearle, P. (1993). Ways to describe dynamical state-vector reduction. Physical Review A, 48(2),
913–923.
Pearle, P. (2015). Relativistic dynamical collapse model. Physical Review D, 91, article 105012.
Pitowsky, I. (1989). Quantum probability, quantum logic (Lecture notes in physics, Vol. 321).
Berlin: Springer.
Pitowsky, I. (1994). George Boole’s ‘conditions of possible experience’ and the quantum puzzle.
British Journal for the Philosophy of Science, 45, 95–125.
Pitowsky, I. (2003). Betting on the outcomes of measurements – A Bayesian theory of quantum
probability. Studies in History and Philosophy of Modern Physics, 34(3), 395–414.
Pitowsky, I. (2007). Quantum mechanics as a theory of probability. In W. Demopoulos & I.
Pitowsky (Eds.), Physical theory and its interpretation – Festschrift in honor of Jeffrey Bub
(Western Ontario series in philosophy of science, Vol. 72, pp. 213–240). New York: Springer.
Pitowsky, I. (2012). Typicality and the role of the Lebesgue measure in statistical mechanics. In Y.
Ben-Menahem & M. Hemmo (Eds.), Probability in physics (pp. 41–58). Heidelberg: Springer.
Popper, K. R. (1982). Quantum theory and the schism in physics: Postscript to the logic of scientific
discovery, Vol. III. W. W. Bartley III (Ed.). London: Hutchinson.
Price, H. (1997). Time’s arrow and archimedes’ point – New directions for the physics of time.
Oxford: Oxford University Press.
Price, H. (2010). Decisions, decisions, decisions – Can Savage salvage Everettian probability? In
S. Saunders, J. Barrett, A. Kent, & D. Wallace (Eds.), Many worlds? Everett, quantum theory,
and reality (pp. 369–391). Oxford: Oxford University Press.
Pusey, M. F., Barrett, J., & Rudolph, T. (2012). On the reality of the quantum state. Nature Physics,
8(6), 475–478.
Rijken, S. (2018). Spacelike and timelike non-locality. Masters Thesis, History and Philosophy of
Science, Utrecht University. https://dspace.library.uu.nl/handle/1874/376363.
Saunders, S. (2002). How relativity contradicts presentism. Royal Institute of Philosophy Supple-
ments, 50, 277–292.
Saunders, S., Barrett, J., Kent, A., & Wallace, D. (Eds.). (2010). Many worlds? Everett, quantum
theory, and reality. Oxford: Oxford University Press.
Sober, E. (1993). Temporally oriented laws. Synthese, 94, 171–189.
3 Unscrambling Subjective and Epistemic Probabilities 89

Uffink, J. (2007). Compendium of the foundations of classical statistical physics. In J. Butterfield &
J. Earman (Eds.), Handbook of the philosophy of physics: Part B (pp. 923–1074). Amsterdam:
North-Holland.
Valentini, A. (1996). Pilot-wave theory of fields, gravitation and cosmology. In J. T. Cushing, A.
Fine, & S. Goldstein (Eds.), Bohmian mechanics and quantum theory – An appraisal (pp. 45–
66). Dordrecht: Springer.
Wallstrom, T. C. (1989). On the derivation of the Schrödinger equation from stochastic mechanics.
Foundations of Physics Letters, 2(2), 113–126.
Watanabe, S. (1965). Conditional probability in physics. Progress of Theoretical Physics Supple-
ment, E65, 135–167.
Wharton, K. (2014). Quantum states as ordinary information. Information, 5, 190–208.
Wallace, D. (2010). How to prove the Born rule. In S. Saunders, J. Barrett, A. Kent, & D. Wallace
(Eds.), Many worlds? Everett, quantum theory, and reality (pp. 227–263). Oxford: Oxford
University Press.
Wallace, D. (2011). The Logic of the past hypothesis. PhilSci preprint http://philsci-archive.pitt.
edu/8894/.
Chapter 4
Wigner’s Friend as a Rational Agent

Veronika Baumann and Časlav Brukner

Abstract In a joint paper Jeff Bub and Itamar Pitowsky argued that the quantum
state represents “the credence function of a rational agent [. . . ] who is updating
probabilities on the basis of events that occur”. In the famous thought experiment
designed by Wigner, Wigner’s friend performs a measurement in an isolated
laboratory which in turn is measured by Wigner. Here we consider Wigner’s friend
as a rational agent and ask what her “credence function” is. We find experimental
situations in which the friend can convince herself that updating the probabilities
on the basis of events that happen solely inside her laboratory is not rational and
that conditioning needs to be extended to the information that is available outside of
her laboratory. Since the latter can be transmitted into her laboratory, we conclude
that the friend is entitled to employ Wigner’s perspective on quantum theory when
making predictions about the measurements performed on the entire laboratory, in
addition to her own perspective, when making predictions about the measurements
performed inside the laboratory.

Keywords Wigner’s friend · (Observer) paradox · Measurement problem ·

Interpretations of quantum theory · Realism · Subjectivism

V. Baumann ()
Vienna Center for Quantum Science and Technology (VCQ), Faculty of Physics, University of
Vienna, Vienna, Austria
Faculty of Informatics, Università della Svizzera italiana, Lugano, Switzerland
e-mail: baumav@usi.ch
Č. Brukner
Vienna Center for Quantum Science and Technology (VCQ), Faculty of Physics, University of
Vienna, Vienna, Austria
Institute of Quantum Optics and Quantum Information (IQOQI), Austrian Academy of Sciences,
Vienna, Austria
e-mail: caslav.brukner@univie.ac.at

© Springer Nature Switzerland AG 2020 91

M. Hemmo, O. Shenker (eds.), Quantum, Probability, Logic,
Jerusalem Studies in Philosophy and History of Science,
https://doi.org/10.1007/978-3-030-34316-3_4
92 V. Baumann and Č. Brukner

4.1 Introduction

Wigner’s-friend-type experiments comprise observations of observers who them-

selves have performed measurements on other systems. In the standard version of
the thought experiment (Wigner 1963) there is an observer called Wigner’s friend
measuring a quantum system in a laboratory, and a so-called superobserver Wigner,
to whom the laboratory including the friend constitutes one joint quantum system.
From the point of view of the friend, the measurement produces a definite outcome,
while for Wigner the friend and the system undergo a unitary evolution. The
discussion surrounding the Wigner’s-friend gedankenexperiment revolves around
the questions: Has the state of the physical system collapsed in the experiment? Has
it collapsed only for the friend or also for Wigner?
Recently, extended versions of Wigner’s-friend scenarios have been devised in
order to show the mutual incompatibility of a certain set of (seemingly) plausible
assumptions. In Brukner (2015, 2018) a combination of two spatially distant
Wigner’s-friend setups is used as an argument against the idea of “observer-
independent facts” via the violation of a Bell-inequality. In Frauchiger and Renner
(2018) two standard Wigner’s-friend setups are combined such that one arrives at
a contradiction when combining the supposed knowledge and reasoning of all the
agents involved. Both arguments are closely related to the quantum measurement
problem and attempted resolutions involve formal as well as interpretational
aspects (Brukner 2015; Baumann and Wolf 2018). At the core of the formal aspects
is the standard measurement-update rule. Without further ontological commitment
regarding a “collapse of the wave function” the latter can be considered to be
the rational-update-belief rule for agents acting in laboratories. The updated state
is the one that gives the correct probabilities for subsequent measurements and
would, hence, be used by a rational agent to make her predictions. That a subjective
(observer dependent) application of the measurement-update rule would lead to the
friend and Wigner making different predictions was already discussed in Hagar
and Hemmo (2006). In general, however, this would be admissible as long as
these predictions cannot directly be tested against each other, i.e. have no factual
relevance.
Here we find a situation in which the friend can verify that her prior prediction
about future measurement outcomes – which is based on applying the standard
state-update rule on the basis of the events that happen inside her laboratory – is
incorrect. In this way, the friend can convince herself, that her prediction cannot be
conditioned solely on the outcomes registered inside her laboratory, but needs to
take into account the information that is available outside of the laboratory as well.
The latter can be transmitted to her by Wigner or an automatic machine from outside
her laboratory. Equipped with this additional information the friend is entitled to
adopt both Wigner’s and her own perspective on the use of quantum theory.
4 Wigner’s Friend as a Rational Agent 93

4.2 Wigner, Friend and Contradictions

We first consider a simple version of our Wigner’s-friend experiment, see also

Finkelstein (2005), moving later to a more
√ general case. Let a source emit a qubit S
(spin-1/2 particle) in state |φS = 1/ 2(| ↑S + | ↓S ), where | ↑S is identified
with “spin up” and | ↓S with “spin down” along the z-direction. The friend F
measures the particle in the σ z -basis (Pauli z-spin operator) with the projectors
MF : {| ↑↑ |S , | ↓↓ |S }. To Wigner W , who is outside and describes the
friend’s laboratory as a quantum system, the measurement of the qubit by the friend
constitutes a unitary evolution entangling certain degrees of freedom of the friend –
her memory that is initially in state |0 (i.e. the register in state “no measurement”)
– with the qubit. Suppose that this results in the overall state

1
|φS |0F → √ (| ↑S |U F + | ↓S |DF ) =: |Φ + SF , (4.1)
2

where |ZF with Z ∈ {U, D} is the state of the friend having registered outcome
z ∈ {“up”, “down”} respectively. Wigner can perform a highly degenerate projective
measurement MW : {|Φ + Φ + |F S , 1 − |Φ + Φ + |F S } to verify his state assignment.
The respective results of the measurement will occur with probabilities p(+) = 1
and p(−) = 0. We note that the specific relative phase between the two amplitudes
in Eq. (4.1) (there chosen to be zero) is determined by the interaction Hamiltonian
between the friend and the system, which models the measurement and is assumed
to be known to and in control of Wigner.
We next introduce a protocol through which Wigner’s friend will realise that her
predictions will be incorrect, if she bases them on the state-update rule conditioned
only on the outcomes observed inside her laboratory. The protocol consists of three
steps (see Fig. 4.1): (a) The friend performs measurement MF and registers an
outcome; (b) The friend makes a prediction about the outcome of measurement

Fig. 4.1 Three steps of the protocol: (a) Wigner’s friend (F) performs a measurement MF on the
system S in the laboratory and registers an outcome. (b) The friend makes a prediction about the
outcome of a future measurement MW on the laboratory (including the system and her memory),
for example, by writing it down on a piece of paper A. She communicates her prediction to Wigner
(W). (c) Wigner performs measurement MW . The three steps (a–c) are repeated until sufficient
statistics for measurement MW is collected. The statistics is noted, for example, on another piece
of paper B. Afterwards, the “lab is opened” and the friend can compare the two lists A and B. An
alternative protocol is given in the main text
94 V. Baumann and Č. Brukner

MW and communicates it to Wigner; (c) Wigner performs measurement MW

Subsequently, Wigner “opens the lab” and the friend can exchange information with
him. Note that apart from the outcome of measurement MF no further information
is given to the friend.
Step (a): Following the laboratory practice and textbook quantum mechanics,
when the friend observers outcome z, she applies the measurement-update rule, and
consequently predicts the probabilities: p(+|z) = |z|Z|Φ + SF |2 = 12 = p(−|z),
where |zS |ZF can be either | ↑S |U F or | ↓S |DF depending on which outcome
the friend observes. It is important to note that the friend’s prediction of the result
of MW does not reveal the result of MF , because for either outcome z the predicted
conditional probability is the same, i.e. p(+) = p(−) = 12 . This means that the
friend’s prediction could be sent out of the laboratory and communicated to Wigner
without changing the superposition state |Φ + SF .
Note that the predictions made by the friend in Wigner’s-friend-type experiments
require her to assign a state to herself, which is sometimes regarded as problematic.
One might reject this possibility altogether and, therefore, claim quantum theory
does not allow the friend to make predictions about Wigner’s measurement results.
Nonetheless, we will see that there is an operationally well-defined procedure,
which if followed by the friend, would give a correct prediction for the measure-
ment. Moreover, since the measurement bases of both the friend and Wigner are
fixed, it is always possible to have a classical joint probability distribution for
their outcomes, whose marginals reproduce the observed distributions by the friend
and Wigner. However, such a classical description of the experiment is no longer
possible if one extends the protocol into one with more measurement choices and
Bell-type set-ups (Brukner 2015, 2018).
Step (b): The friend opens the laboratory in a manner that allows a record of
her prediction (e.g., a specific message A written on a piece of paper) to be passed
outside to Wigner, keeping all other degrees of freedom fully isolated and in this
way preserving the coherence of the state. The piece of paper waits outside of the
laboratory to be evaluated later in the protocol. The state that Wigner assigns to the
friend’s laboratory at the present stage is |Φ + SF |p(+) = p(−) = 1/2A , where the
second ket refers to the message.
Step (c): After the friend communicates her prediction to the outside, Wigner
(or an automatic machine) performs measurement MW and records the result, for
example on a separate piece of paper B. At this stage of the experiment, Wigner
assigns to the friend’s laboratory and the two messages the state |Φ + SF |p(+) =
p(−) = 1/2A |p(+) = 1, p(−) = 0B with the two last kets representing the two
lists.
We note that the measured state |Φ + SF is an eigenstate of the measured
operator. This implies that outcome “+” will occur with unit probability, and
furthermore that the measurement does not change the state of the laboratory.
Hence the three steps (a–c) of the protocol can be repeated again and again without
changing the initial superposition state until a sufficient statistics in measurement
MW is collected. As a result, the two pieces of paper, one with the friend’s prediction
4 Wigner’s Friend as a Rational Agent 95

A = {p(+) = 12 , p(−) = 12 }, and one with the actually observed relative

number of counts for the two outcomes B = {p(+) = 1, p(−) = 0}, display a
statistically significant difference. In the very last step of the protocol ”the laboratory
is opened” (which is equivalent to Wigner performing a measurement in the basis
{|zz|S ⊗ |Z  Z  |F }), such that the joint states for the systems and the friend’s
memory are reduced to the form |zS |ZF . Irrespective of the specific state to which
the friend is reduced, she can now compare the two messages and convince herself
that her prior prediction deviates from the actually observed statistics.
In an alternative protocol the friend resides in the laboratory for the duration of
the whole protocol. Steps (a) and (c) are retained, but instead of step (b), the friend
either receives the result of measurement MW from Wigner after each measurement
run or the entire list B at the end of the protocol. In either case the friend can
compare the two lists inside the laboratory and arrive at the same conclusion as
in the previous protocol. One may object that in the present case the friend’s
conclusion relies on Wigner being trustful and reliable when providing her with
the measurement result of MW . As mentioned above, this is irrelevant as Wigner
can be replaced in the protocol by an automatic machine that the friend could have
pre-programmed herself.
The discrepancy between the friend’s prediction and the actual statistics can
be made as high as possible in a limit. Consider that  the source emits a higher
d
dimensional quantum system in state |φ d S = √1 j =1 |j S and the friend
d
measures in the respective basis, MF : {|j j |S } with j = 1 . . . d. Wigner describes
the measurement of the friend as a unitary process that results in state

1 
d
|φ d S |0F → √ |j S |α j F =: |Φd+ SF , (4.2)
d j =1

where |α j F is the state of the friend’s memory after having observed outcome
j . For MW we choose: {|Φd+ Φd+ |SF , 1 − |Φd+ Φd+ |SF }. Wigner records the “+”
result with unit probability, i.e. p(+) = 1 and p(−) = 0. Using the measurement-
update rule the friend, however, predicts

1
p(+|j ) = |j |α j |Φd+ SF |2 = −→ 0
d d→∞
(4.3)
1
p(−|j ) = 1 − −→ 1,
d d→∞

independently of the actual outcome she registers in measurement MF .

Either the three steps (a–c) of the protocol in Fig. 4.1 can be repeatedly applied
or Winger and the friend follow the alternative protocol. In both cases, the friend
can convince herself, without relying on the knowledge other agents may have,
that she made an incorrect prediction of future statistics using the state-update rule
conditioned solely on the observations made within her laboratory. As the dimension
96 V. Baumann and Č. Brukner

of the measured system and the number of outcomes increases, the discrepancy
between her prediction and the actual statistics becomes maximal, giving rise to an
“all-versus-nothing” argument.

4.3 Discussion

We now discuss possible responses of different interpretations and modifications of

quantum theory to the experimental situation described above.
(1) Theories that deny universal validity of quantum mechanics. The theories
postulate a modification of standard quantum mechanics, such as sponta-
neous (Ghirardi et al. 1986) and gravity-induced (Penrose 2000; Diósi 2014)
collapse models, which become significant at the macroscopic scale. These
models could exclude superpositions of observers as required in Wigner’s-
friend-type experiments and hence the protocol cannot be run. In our view this
is the most radical position as it would give rise to new physics.
(2) Many-worlds interpretation (Wheeler 1957). These incorrect prediction of the
friend is due to her lack of knowledge of the total state of the laboratory
including the friend’s memory (i.e. “the wave function of the many-worlds”).
If the friend would know this state, she could make a prediction that is in
accordance with actually observed statistics.
(3) Quantum Bayesian (Fuchs 2010), neo-Copenhagen (Brukner 2015), and rela-
tional interpretation (Rovelli 1996). The quantum state is always given relative
to an observer and represents her/his knowledge or degree of belief. It is natural
for an agent to update his or her degree of belief in face of new information. As
a consequence of newly acquired knowledge in the protocol the friend would
update her degree of belief and assign a new state to the laboratory that is in
agreement with the observations.
Note that in both interpretations (2) and (3) the friend can learn the state of
the laboratory as described by Wigner. Wigner can simply communicate either the
initial state |φS |0F of the system together with the Hamiltonian of the laboratory,
or the final state |Φd+ SF of the laboratory to the friend before step (b) begins.
The important feature is that this can be done without destroying the coherence
of the quantum state of the laboratory. In this way, the friend can learn the overall
state of her laboratory from the perspective of an outside observer and include it in
making her prediction. As a consequence, the friend may operate with a pair of states
{|zS , |Φd+ SF }. The first component is used for predictions of measurements made
on the system alone and is conditional on the registered outcome in the laboratory;
the second component is used for predictions of measurements performed on the
entire laboratory (= “system + friend”). When an outcome of a measurement on the
system is registered in the laboratory, the friend applies the state-update rule on the
first component, without affecting the second component of the pair of states.
4 Wigner’s Friend as a Rational Agent 97

Alternatively, the friend can use only the quantum state Wigner assigns to the
laboratory, but apply a modification of the Born rule as introduced in Baumann
and Wolf (2018). The new rule enables to evaluate conditional probabilities in a
sequence of measurements. In the case when the sequence of measurements is
performed on the same system, it restores the predictions of the standard state-
update rule. However, the rule enables making predictions in Wigner’s-friend
scenarios as well, where the first measurement is performed on the system by
Wigner’s friend, and the subsequent measurement is performed on the “system +
friend” by Wigner. In this case, the rule recovers the prediction as given by
Eq. (4.3). This modified Born rule gains an operational meaning in the present
setup where Wigner’s friend has access to outcomes of both measurements and can
evaluate the computed conditional probabilities (Note that Wigner does not have
this status, since he has no access to the outcome of the friend’s measurement.).
While Bub and Pitowsky claim in Bub and Pitowsky (2010) that “Conditionalizing
on a measurement outcome leads to a nonclassical updating of the credence
function represented by the quantum state via the von Neumann-Lüders rule. . . ” we
argue that in case of Wigner’s friend experiments both the “outside” and “inside”
perspective should contribute to “the credence function” for making predictions.

Modified Born Rule (Baumann and Wolf 2018)

We apply the modified Born rule of Baumann and Wolf (2018) to the present situa-
tion. The rule defines the conditional probabilities in a sequence of measurements.
According to it, every measurement is described as a unitary evolution analogous
to Eq. (4.1). Consider a sequence of two measurements, the first one described by
unitary U and the subsequent one by unitary V . The sequence of unitaries correlate
the results of the two measurements with two memory registers 1 and 2. The overall
state of the system and the two registers evolve as

U 
|φS |01 |02 −
→ j |φ|j S |α j 1 |02 (4.4)
j

V 
−
→ j |φβ̃ j k |j, α j |β̃ j k S1 |β k 2 = |Φtot .
jk

The conditional probability for observing outcome k̄ given that outcome j¯ has been
observed in the previous measurement is given by

p(k̄, j¯)
p(k̄|j¯) =  . (4.5)
¯
k̄ p(k̄, j )
98 V. Baumann and Č. Brukner

Here p(k̄, j¯) is the joint probability for observing the two outcomes and is given
by the projection of the overall state |Φtot  on the states |α j¯ 1 and |β k̄ 2 of the two
registers:

¯ k̄) = Tr[(1S ⊗ |α ¯ α ¯ |1 ⊗ |β β |2 )|Φtot Φtot |]

p(j, j j k̄ k̄
(4.6)

In the case when the sequence of the measurements are performed on a single
system, the rule restores the prediction of the standard state-update rule, i.e.
p(k̄, j¯) = |k̄|j¯|2 . This corresponds to the state

|Φtot  = j |φk|j |kS |α j 1 |β k 2 . (4.7)
jk

For Wigner’s-friend scenarios, however, the friend would predict conditional

probabilities for Wigner’s result + given her result j that are in accordance with
Wigner’s observations Eq. (4.3). In this case the overall state is


d
|Φtot  = j |φ|j S |α j 1 |+2 (4.8)
j =1

where |+2 is the state of the memory register corresponding to observing the result
of |Φd+ Φd+ |S1 , and the joint probability is given by

p(+, j ) = Tr[(1S ⊗ |α j α j |1 ⊗ |++|2 )|Φtot Φtot |]. (4.9)

This means that full unitary quantum theory together with the adapted Born rule
allows for consistent predictions in Wigner’s-friend-type scenarios, while giving the
same predictions as the standard Born and measurement-update rules for standard
observations.

Acknowledgements We thank Borivoje Dakic, Fabio Costa, Flavio del Santo, Renato Rennner
and Ruediger Schack for useful discussion. We also thank Jeff Bub for the discussions and for
the idea of the alternative protocol. We acknowledge the support of the Austrian Science Fund
(FWF) through the Doctoral Programme CoQuS (Project no. W1210-N25) and the projects I-
2526-N27 and I-2906. This work was funded by a grant from the Foundational Questions Institute
(FQXi) Fund. The publication was made possible through the support of a grant from the John
Templeton Foundation. The opinions expressed in this publication are those of the authors and do
not necessarily reflect the views of the John Templeton Foundation.
4 Wigner’s Friend as a Rational Agent 99

References

Baumann, V., & Wolf, S. (2018). On formalisms and interpretations. Quantum, 2, 99.
Brukner, Č. (2015). On the quantum measurement problem. arXiv:150705255. https://arxiv.org/
abs/1507.05255.
Brukner, Č. (2018). A no-go theorem for observer-independent facts. Entropy, 20(5), 350.
Bub, J., & Pitowsky, I. (2010). Two dogmas about quantum mechanics. In Many worlds?: Everett,
quantum theory, & reality (pp. 433–459). Oxford: Oxford University Press.
Diósi, L. (2014). Gravity-related wave function collapse. Foundations of Physics, 44(5), 483–491.
Finkelstein, J. (2005). “quantum-states-as-information” meets Wigner’s friend: a comment on
hagar and hemmo. arXiv preprint quant-ph/0512203.
Frauchiger, D., & Renner, R. (2018). Quantum theory cannot consistently describe the use of itself.
Nature Communications, 9(1), 3711.
Fuchs, C. A. (2010). Qbism, the perimeter of quantum bayesianism. arXiv:10035209. https://arxiv.
org/abs/1003.5209.
Ghirardi, G. C., Rimini, A., & Weber, T. (1986). Unified dynamics for microscopic and macro-
scopic systems. Physical Review D, 34(2):470. https://doi.org/10.1103/PhysRevD.34.470.
Hagar, A., & Hemmo, M. (2006). Explaining the unobserved—Why quantum mechanics ain’t only
about information. Foundations of Physics, 36(9), 1295–1324.
Penrose, R. (2000). Wavefunction collapse as a real gravitational effect. In Mathematical physics
2000 (pp. 266–282). Singapore: World Scientific.
Rovelli, C. (1996). Relational quantum mechanics. International Journal of Theoretical Physics,
35(8), 1637–1678.
Wheeler, J. A. (1957). Assessment of everett’s “relative state” formulation of quantum theory.
Reviews of Modern Physics, 29(3), 463.
Wigner, E. P. (1963). The problem of measurement. American Journal of Physics, 31(1), 6–15.
https://doi.org/10.1119/1.1969254.
Chapter 5
Pitowsky’s Epistemic Interpretation
of Quantum Mechanics and the PBR
Theorem

Yemima Ben-Menahem

Abstract In his work on the foundations of Quantum Mechanics, Itamar Pitowsky

set forth an epistemic interpretation of quantum state functions. Shortly after his
untimely death, Pusey, Barrett and Rudolph published a theorem—now known as
the PBR theorem—that purports to rule out epistemic interpretations of quantum
states and render their realist interpretations mandatory. If one accepts this conclu-
sion, the theorem is fatal to Pitowsky’s interpretation. The aim of this paper is to cast
doubt on this verdict and show that Pitowsky’s position is actually unaffected by the
PBR theorem. Had it been a form of instrumentalism, the defense of Pitowsky’s
view would have been trivial, for according to its authors, the PBR theorem does
not apply to instrumentalist readings of quantum mechanics. I argue, however, that
Pitowsky’s position differs significantly from instrumentalism, and yet, remains
viable in the face of the PBR theorem.

Keywords Realism · Instrumentalism · Probability · Subjective probability ·

Quantum mechanics · PBR theorem · Epistemic interpretation

5.1 Introduction

The probabilistic interpretation of Schrödinger’s equation, offered by Max Born

in 1926, constitutes as essential component of standard interpretations of quantum
mechanics (QM).1 In comparison with the use of probability in physics up to that

This paper draws on an earlier paper on the PBR theorem (Ben-Menahem 2017).

1 Standardinterpretations are mostly descendants of the Copenhagen interpretation to which Born

contributed and subscribed. Standard interpretations need not embrace such Copenhagen tenets

Y. Ben-Menahem ()
Department of Philosophy, The Hebrew University of Jerusalem, Jerusalem, Israel
e-mail: yemima.ben-menahem@mail.huji.ac.il

© Springer Nature Switzerland AG 2020 101

M. Hemmo, O. Shenker (eds.), Quantum, Probability, Logic,
Jerusalem Studies in Philosophy and History of Science,
https://doi.org/10.1007/978-3-030-34316-3_5
102 Y. Ben-Menahem

point, the novelty of Born’s interpretation was that it construed quantum probabili-
ties as fundamental in the sense that they were believed to be irreducible, that is, they
were not alleged to arise from averaging over an underlying deterministic theory. In
other words, quantum probabilities were said to be different from those of other
probabilistic theories, statistical mechanics in particular, where the existence of an
underlying deterministic theory—classical mechanics—is taken for granted. The
proclaimed irreducibility gave rise to serious conceptual puzzles, (more on these
below), but, from the pragmatic/empirical point of view, the notion that quantum
states (as represented and calculated by QM) stand for probability amplitudes of
measurement results was sufficient to make quantum mechanics a workable (and
highly successful) theory.
Itamar Pitowsky endorsed the irreducibly probabilistic nature of QM, adding to
it a major twist. On the basis of developments such as Bell’s inequalities, Gleason’s
theorem and other no-go theorems, he argued that QM is not merely a probabilistic
theory, but a non-classical probabilistic theory.2 Studying the structure of the
new theory in both algebraic and geometrical terms, he was able to demonstrate
that quantum probabilities deviate from classical probabilities in allowing more
correlations between events than classical probability theory would allow. He saw
his interpretation as a response to the threat of non-locality and as a key to the
solution of other puzzles such as the notorious measurement problem. In addition,
his understanding of QM as a theory of non-classical probability also constitutes
a new take on the status of the theory of probability, turning it from an a priori
extension of logic into an empirical theory, which may or may not apply to particular
empirical domains.3
The meaning of the concept of probability, which has been debated among
scholars for a long time, is of particular concern in the context of QM. An option
that has attracted a great deal of attention in the last few decades is an epistemic
interpretation of probability (also known as information-theoretic, or Bayesian),
according to which probabilities in general, and quantum probabilities in particular,
represent states of knowledge or belief or information. How are we to complete
this sentence—knowledge (belief, information) about what? It turns out that the

as complementarity (which had actually been debated by members of the Copenhagen school),
but from the perspective of this paper and the problems about probability it discusses, the term
‘standard interpretation’ will do. I will not discuss the notion of probability in non-standard
approaches such as Bohm’s theory or the Many Worlds interpretation.
2 The non-classical character of quantum probability had been recognized, for example, by Richard

Feynman: “But far more fundamental was the discovery that in nature the laws of combining
probabilities were not those of the classical probability theory of Laplace” (Feynman 1951, p.
533). In actually developing and axiomatizing the non-classical calculus of quantum probability,
Pitowsky went far beyond that general recognition. I will not undertake a comparison of Pitowsky’s
work with that of his predecessors or contemporaries.
3 As a theory of rational belief, the theory of probability can be seen as an extension of logic

and the idea that the theory of probability is empirical—as extension of the view that logic is
empirical. Pitowsky’s interpretation of QM is thus related to the earlier suggestion that QM obeys
a non-classical logic—quantum logic. The connection is explicit in the title of Pitowsky’s (1989)
Quanyum Probability—Quantum Logic.
5 Pitowsky’s Epistemic Interpretation of Quantum Mechanics and the PBR Theorem 103

question is far from trivial; theorists who share an epistemic interpretation of

quantum mechanics may still differ significantly regarding its answer.4 As Pitowsky
subscribed to an epistemic interpretation of quantum probabilities, his answer is
central to what follows and will be examined in detail.
In 2011, a year after Pitowsky’s premature death, Pusey, Barrett and Rudolph
published the theorem now known as the PBR theorem (Pusey et al. 2012), which
purports to rule out epistemic interpretations of QM and vindicate realist ones.5
Pitowsky’s epistemic interpretation thus appeared to be undermined. I will argue,
however, that it is not. It should be stressed right away that Pitowsky’s interpretation
is not a form of instrumentalism. If it were, the argument would be trivial, for—as its
authors concede—the PBR theorem does not apply to instrumentalist understand-
ings of quantum mechanics. Exploring the difference between Pitowsky’s position
and instrumentalism in terms of their empirical implications, I will show that despite
its departure from instrumentalism, Pitowsky’s interpretation, remains viable in the
face of the PBR theorem. The structure of the paper is the following: In Sect. 5.2,
I describe some of the above-mentioned conceptual problems that, from early on,
plagued the probabilistic interpretation of QM. Section 5.3 provides a summary of
Pitowsky’s epistemic interpretation and Sect. 5.4—a summary of the PBR theorem.
In Sect. 5.5, I explain why Pitowsky’s interpretation is not vulnerable to the PBR
theorem. Section 5.6 addresses the thorny issue of instrumentalism.

5.2 Born’s Probabilistic Interpretation and Its Problems

In 1926, upon realizing that Schrödinger’s wavefunction cannot represent an ordi-

nary three-dimensional wave, Born offered his probabilistic interpretation according
to which quantum states, described by Schrödinger’s Ψ function, represent proba-
bility amplitudes. Born was explicit about the curious character of Ψ under this
interpretation: Although particles follow probabilistic laws, he said, the probability
itself transforms “in accordance with the causal principle,” i.e., deterministically.
Furthermore, QM only answers well-posed statistical questions, and remains silent
about the individual process. Born therefore characterized QM as a “peculiar blend
of mechanics and statistics,”—“eine eigenartige Verschmelzung von Mechanik und
Statistik”.6 This formulation is almost identical to that given by Jaynes and quoted
in the PBR paper

4 Priorto the publication of the PBR theorem, the difference between possible answers to this
question was rather obscure. One of the contributions of the theorem is that it sheds light on this
difference.
5 The 2011 version of the paper was published in the arXiv quant-ph11113328.
6 Born [1926] (1963), p. 234. “Die Bewegung der partikeln folgt Wharscheinlichkeitsgesetzen,

die wharscheinlichkeit selbst aber breitet sich im Einklang mit dem Kausalgesetz aus.” . . . Die
Quantenmechanik allenfals nur Antwort gibt auf richtig gestellte statistische Fragen, aber im
allgemeinen die Frage nach dem Ablauf sines Einzelprozesses unbeantwortet laesst.”
104 Y. Ben-Menahem

Our present [quantum mechanical] formalism . . . is a peculiar mixture describing in part

realities of Nature, in part incomplete human information about Nature, all scrambled up
by Heisenberg and Bohr into an omelette that nobody has seen how to unscramble. (Pusey
et al. 2012, p. 475).7

The majority of quantum physicists were willing to tolerate the paradoxical

situation diagnosed by Born. They treated QM both as a probabilistic theory and
as a fundamental physical theory in which the wave equation plays the role of the
equations of motion in classical mechanics. The question of whether the concept of
probability could sustain this Verschmelzung was left hanging in midair. A dissident
minority, including Einstein, preferred to bite the bullet and understand QM as a full-
blown probabilistic theory, that is, a statistical description of an ensemble of similar
systems that makes no definite claims about individual members of the ensemble.8
Hence, the immediate analogy with statistical mechanics.
Statistical mechanics does indeed hold promise for peaceful coexistence between
the probabilistic description of an ensemble and the precise physical descrip-
tion of each individual system comprising the ensemble.9 Even if this precise
description is inaccessible in practice, it is taken to be conceivable in principle.
The analogy between statistical and quantum mechanics suggested that quantum
states correspond to macrostates in statistical mechanics, implying that in QM too,
individual systems could occupy determinate physical states. These determinate
states, though, like microstates in statistical mechanics, are not directly represented
by quantum states; the latter only represent the statistical behavior of the ensemble.
The picture of an underlying fundamental level, the description of which bears to
QM the relation that classical mechanics bears to statistical mechanics, seemed
to diminish the novelty and uniqueness of QM. It was only natural, then, that
advocates of the Copenhagen interpretation opposed this picture and the analogy
with statistical mechanics on which it was based. For them, QM was a novel
theory—complete, fundamental and irreducible—that replaces classical mechanics
rather than supplementing it with a convenient way of treating multi-particle
systems. They therefore disagreed with ensemble theorists about the prospects of
completing QM by means of ‘hidden variable’ theories.

7 Jaynes presupposes an epistemic interpretation of probability, but the scrambling he refers to

does not depend on it. We could replace the epistemic construal of probability with an objective
ensemble interpretation and the question would remain the same: What, exactly, does the state
function represent? Nonetheless, there are significant differences between the ensemble and
epistemic interpretations, which the PBR theorem helps clarify, and which will be addressed later.
8 The ensemble interpretation was recommended, for example by Blokhintsev (1968), Ballentine

(1970), Popper (1967).

9 It is well-known that there are serious problems regarding the compatibility of statistical

mechanics and classical mechanics, but these pertain mainly to the issue of irreversibility and will
not concern us here. See Albert (2000), Hemmo and Shenker (2012), and the literature there cited.
Note, further, that despite the fact that the central laws of statistical mechanics are probabilistic,
there is no binding reason to construe these probabilities as subjective.
5 Pitowsky’s Epistemic Interpretation of Quantum Mechanics and the PBR Theorem 105

Neither of these competing interpretations of quantum probabilities was epis-

temic. In this respect, there is a significant difference between the alternatives
debated in the first decades of QM and the contemporary alternatives juxtaposed
in the PBR argument, where the divide is between objective (physical, ontic)
and epistemic interpretations. Given that epistemic interpretations of probability
in general (outside of the quantum context) came of age later than objective
interpretations, this difference should not surprise us.10 It is significant, though,
that when, in the wake of work by de Finetti and Savage, epistemic interpretations
became popular, they were understood as diverging from traditional interpretations
only in their conceptual basis, not in the probabilities calculated by them. Indeed, it
was considered to be the great achievement of these epistemic interpretations that
despite their subjective starting point, they recovered the rules of objective theories
of probability and proved to be equivalent to them.
The equivalence between subjective and objective interpretations of the concept
of probability holds for most applications of probability theory. In ordinary contexts,
such as gambling or insurance, one could think of probabilities as reflecting our
ignorance about the actual result rather than in terms of frequencies, ensembles
or propensities, and it wouldn’t matter. The equivalence also holds in statisti-
cal mechanics. In the above quotation from Jaynes, for instance, he adopts an
epistemic interpretation of statistical mechanical probabilities, but this subjective
interpretation does not yield different predictions than those of standard statistical
mechanics. The epistemic interpretation of QM, however, presumes to make an
empirical difference—this, as we will see, is the crux of the PBR theorem. Bearing
this difference in mind (and returning to it later), let us probe a bit deeper into
the controversy between the majority view that swallowed Born’s Verschmelzung
and the dissidents promoting the ensemble interpretation. The focal question on
which the parties were divided was whether quantum states tell us anything about
the individual system. Here are some of the pros and cons of the conflicting positions
on this question.
The great merit of the ensemble interpretation was its readymade response to
the problem regarding the ‘collapse’ of the wavefunction. If the wavefunction
represents the physical state of an individual system, its instantaneous ‘collapse’ (to
a sharp value) upon measurement is worrisome. As a physical process, it should
be restricted by the known constraints on physical processes, in particular, the
spatiotemporal constraints of the Special Theory of Relativity (STR). But once
quantum probabilities are assumed to refer to an ensemble of similar systems,
the fact that measurements on individual systems yield definite values is exactly
what one would expect! The quantum collapse, under this interpretation, is no more
troubling than the ‘collapse’ of a flipped coin from a state of probability 0.5, which

10 The epistemic interpretation already existed in 1930 but took some time to make an impact
outside of the community of probability theorists. Leading advocates of the subjective/epistemic
interpretation of probability as degree of belief are Ramsey (1931), de Finetti and Savage (1954).
The delay in impact is clearly reflected in the fact that de Finetti’s works of the 1930s had to wait
several decades to be collected and translated into English (de Finetti 1974).
106 Y. Ben-Menahem

characterizes an ensemble, to either heads or tails when a single coin is flipped. In

other words, since, on the ensemble interpretation, the Ψ function does not stand for
a physical state of an individual system in the first place, worries about the physical
properties of the collapse, its Lorentz invariance etc., are totally misplaced.
Despite this advantage, the ensemble interpretation also faces a number of serious
questions.
(a) Interference and other periodic effects. What do quantum phenomena such
as superposition, interference and entanglement mean under an ensemble
interpretation? Superposition and interference had an intuitive meaning
in Schrödinger’s original picture of wave mechanics, but he too became
disillusioned when realizing that the wave function was a multi-dimensional
wave in configuration space, not an ordinary wave in three-dimensional space.
Born’s interpretation of Ψ as representing a probability amplitude had the
advantage that Ψ no longer needed to be situated in three dimensional space,
and the disadvantage that the meaning of superposition and interference became
mysterious. Getting a pattern on the screen from the interference of abstract
probability waves has been compared to getting soaked from the probability of
rain.11 One could also formulate this question in terms of the meaning of the
phase of the wave function.12 Periodicity appears in all versions of QM, but the
expression representing this periodicity–the phase–disappears in the calculation
of probabilities. If we are only interested in probabilities and deny the reality
of the wave, there is no good explanation for the significance of the phase.
An intriguing example that suggests a physical meaning of the phase is the
Aharonov-Bohm effect where an interference pattern resulting from a phase-
shift in a field-free region is produced. It is difficult to explain this interference
from the perspective of an ensemble interpretation.13
(b) The uncertainty relations. What is the status of Heisenberg’s uncertainty prin-
ciple from the perspective of the ensemble interpretation? Presumably, the
principle should be construed as a statistical law, which constrains the dis-
persion of values in an ensemble of systems, not the definiteness of values
representing properties of individual systems. It should thus be compatible
with QM for an individual system to have sharp values of all variables
including canonically conjugate ones. In other words, on the ensemble interpre-
tation, the uncertainty relations could be violated! Proponents of the ensemble
interpretation in fact entertained such violation. In a 1935 letter to Einstein,

11 Iam not sure about the origin of this analogy; I may have heard it from David Finkelstein.
12 The meaning of the phase is being examined by Guy Hetzroni in his Ph.D dissertation The
Quantum Phase and Quantum Reality submitted to the Hebrew University in July 2019 and
approved in October 2019. See also Hetzroni 2019a.
13 The Aharonov Bohm paper was published in 1959 but the problem about the phases could have

been raised in 1926. There are different interpretations of the Aharonov Bohm effect (see, for
example Aharonov and Rohrlich 2005, Chaps. 4, 5 and 6 and Healey 2007, Chap. 2), but, none of
them is intuitive from the perspective of an ensemble interpretation.
5 Pitowsky’s Epistemic Interpretation of Quantum Mechanics and the PBR Theorem 107

Popper proposed a thought experiment designed to illustrate a system whose

position and momentum are both well-defined. In his response (written a short
while after the EPR paper; Einstein et al. 1935), Einstein rejected Popper’s
attempt to violate the uncertainty relations directly, but argued that one could,
nonetheless, circumvent them in the split system described in the EPR paper.14
Evidently, Einstein did not see the uncertainty relations as applying to individual
systems; in accordance with the ensemble interpretation, he took the values of
conjugate observables to be well-determined, though not directly measurable.
This statistical understanding of the uncertainly relations was unacceptable
to the Copenhagen school and was later challenged on both theoretical and
empirical grounds.
(c) Disturbance by measurement. Thought experiments by Heisenberg and others
motivated the uncertainty principle by showing how measurement of one
quantum observable, say position, destroys the value of another observable
(measured on the same particle or system), in this case, momentum.15 On
the ensemble interpretation, it appears, this ‘disturbance’ could be avoided,
for this interpretation takes QM to pertain to a multitude of measurements
made on many different particles or many different systems. In that case,
however, the lesson of the convincing thought experiments becomes dubious.
Is it possible to get around disturbance? The Copenhagen school declined this
possibility and treated disturbance as fundamental. Later experiments failed
to eliminate disturbance, thus constituting another difficulty for the ensemble
interpretation.16
With these questions in mind, we can return to the contentious analogy between
QM and statistical mechanics. The ensemble theorist who embraces this analogy, we
now realize, is not only committed to a more fundamental level than the quantum
level (i.e., underlying ‘hidden variables’), but is also at variance with the standard
interpretation about the meaning of central quantum laws such as the uncertainty
relations! It is therefore slightly misleading to use the term ‘interpretation’ in this

14 The letter is reproduced in Popper (1968) pp. 457–464. Incidentally, Einstein’s letter repeats
the main argument of the EPR paper in terms quite similar to those used in the published version
and makes it clear that the meaning of the uncertainty relations is at the center of the argument.
The letter thus disproves the widespread allegations that Einstein was unhappy with the published
argument and that it was not his aim to question the uncertainty relations.
15 It was only in the first few years of QM that Heisenberg and Bohr understood the uncertainty

relations as reflecting concrete disturbance by measurement, that is, as reflecting a causal process.
In his response to the EPR paper, Bohr realized that no concrete disturbance takes place in the EPR
situation. He construed the uncertainty relations, instead, as integral to the formalism, and as such,
holding for entangled states even when separated to spacelike distances so that they cannot exert
any causal influence on one another.
16 In contrast to the older ensemble interpretation, current epistemic interpretations of QM do not

seek to evade disturbance. Rather, disturbance is understood in these interpretations as loss of

information, a quantum characteristic resulting from the underlying assumption of maximal, and
yet less-than-complete, information.
108 Y. Ben-Menahem

context; the two positions actually represent different theories diverging in their
empirical implications.17
Thus far, I have concentrated on the distinction between the ensemble interpreta-
tion and standard QM rather than the distinction between epistemic and ontic models
that is the focus of the PBR theorem. The reason for using the former as backdrop
for the latter is that in general, ensemble and epistemic interpretations have not been
properly distinguished by their proponents and do indeed have much in common;
neither of them construes quantum states as a unique representation of physical
states of individual systems and both can dismiss the worry about the collapse of the
wavefunction. In principle, however, epistemic and ensemble interpretations could
diverge: The epistemic theorist could go radical and decline the assumption of an
underlying level of well-determined physical states that was the raison d’être of the
ensemble interpretation. This was in fact the route taken by Pitowsky.

5.3 Pitowsky’s Interpretation of QM

Pitowsky’s approach to QM is based on three very general insights. First, the

intriguing nature of QM—its departure from classical mechanics—already man-
ifests itself at the surface level of quantum events and their correlations. Thus,
before undertaking a theoretical explanation (in terms superposition, interference,
entanglement, collapse, nonlocality, duality, complementarity, and so on), one must
identify and systematize the observed phenomena. Secondly, classical probability
theory seems to be at odds with various observable quantum phenomena.18 Pitowsky
therefore conjectured that a non-classical probability theory that deviates from its
classical counterpart in its empirical implications provides a better framework for
QM. Thirdly, once we follow this route and construe QM as a non-classical theory
of probability, no further causal or dynamical explanation of quantum peculiarities
is required. Characteristic quantum phenomena such as nonlocality fall into place in
the new framework in the same way that rod contraction (in the direction of motion)
and time dilation fall into place in the spacetime structure of STR. In both cases, the
new framework renders the causal/dynamical explanation redundant.
The non-classical nature of quantum probability manifests itself in the violation
of basic classical constraints on the probabilities of interrelated events. For example,
in the classical theory of probability, it is obvious that for two events E1 and E2
with probabilities p1 and p2 , and an intersection whose probability is p1 .p2 , the

17 Construing the diverging interpretations as different theories does not mean that the existing
empirical tests are conclusive, but that in principle, such tests are feasible. Even the evidence
accumulated so far, however, makes violation of the uncertainty relations in the individual case
highly unlikely.
18 There are of course other approaches to QM, some of which retain classical probability theory.

The ensemble interpretation discussed above and Bohm’s theory, which I do not discuss here are
examples of such approaches.
5 Pitowsky’s Epistemic Interpretation of Quantum Mechanics and the PBR Theorem 109

probability of the union (E1 UE2 ) is p1 + p2 −p12 , and cannot exceed the sum of the
probabilities (p1 + p2 ).

0 ≤ p1 + p2 − p12 ≤ p1 + p2 ≤ 1

The violation of this constraint is already apparent in simple paradigm cases such
as the two-slit experiment, for there are areas on the screen that get more hits when
the two slits are both open for a certain time interval Δt, than when each slit is
open separately for the same interval Δt. In other words, contrary to the classical
principle, we get a higher probability for the union than the sum of the probabilities
of the individual events. (Since we get this violation in different experiments—
different samples—it does not constitute an outright logical contradiction.) This
quantum phenomenon is usually described in terms of interference, superposition,
the wave particle duality, the nonlocal influence of one open slit on a particle passing
through the other, and so on. Pitowsky’s point is that regardless of the explanation of
the pattern predicted by QM (and confirmed by experiment), we must acknowledge
the bizarre phenomenon that it displays—nothing less than a violation of a highly
intuitive principle of the classical theory of probability.
QM predicts analogous violations of other classical conditions. The most famous
of these violations is that of the Bell inequalities which are violated by QM (and
experiment). Pitowsky showed that from the analogue of the above rule for three
events, it is just a short step to Bell’s inequalities. He therefore linked the inequalities
directly to Boole’s “conditions of possible experience” (the basic conditions of
classical probability) and their violation to the need for an alternative probability
calculus. To appreciate the significance of this move for the debate over the
epistemic interpretation, it is crucial to understand that what makes the violated
conditions classical is their underlying assumption of determinate states occupied
by the entities in question (or properties they obtain) independently of measurement:
Just as balls in an urn have definite properties such as being red or wooden, to
derive Bell’s inequalities it is assumed that particles have a definite polarization
or a definite spin in a certain direction, and so on. The violation of the classical
principles of probability compels us to discard this picture, replacing it with a new
understanding of quantum states and quantum properties. Rather than an analogue
of the classical state, which represents physical entities and their properties prior
to measurement, Pitowsky saw the quantum state function as keeping track of the
probabilities of measurement results, “a device for the bookkeeping of probabilities”
(2006, p. 214).19
This understanding of the state function, in turn, led Pitowsky to two further
observations: First, he interpreted the book-keeping picture subjectively, i.e., quan-

19 Pitowsky’s answer is similar to that of Schrödinger, who characterizes quantum states as “the
momentarily-attained sum of theoretically based future expectations, somewhat laid down as in a
catalog. It is the determinacy bridge between measurements and measurements” (1935, p. 158).
Pitowsky only became aware of this similarity around 1990. For more on Schrödinger’s views and
their similarity to those of Pitowsky see note number 25 below and Ben-Menahem (2012).
110 Y. Ben-Menahem

tum probabilities are understood as degrees of partial belief of a rational observer

of quantum phenomena. Second (and here he joins other epistemic interpretations),
the notorious collapse problem is not as formidable as when the state function is
construed realistically, for if what collapses is not a real entity in physical space,
there is no reason for the collapse to abide by the constraints of locality and Lorentz
invariance.20 Consequently, he renounced what he and Bub, in their joint paper
(2010), dub “two dogmas” of the received view, namely the reality of the state
function and the need for a dynamic account of the measurement process.
In his (2006), Pitowsky proposed an axiomatic system for QM that elaborates the
earlier axiomatization by Birkhoff and von Neumann. In particular, he reworked the
relation between the Hilbert space structure of quantum events and projective geom-
etry. Furthermore, he sought to incorporate later developments such as Gleason’s
theorem and identify their connection with the axiom system. The ramifications
of the non-classical structure of the quantum probability space, he argued, include
indeterminism, the loss of information upon measurement, entanglement and Bell-
type inequalities. As Pitowsky and Bub argued in their joint 2010, the implications
of the axiom system are also closely related to the information-theoretic principles
of no-cloning(for pure states) or no broadcasting (for mixed states) according to
which an arbitrary quantum state cannot be copied (Park 1970; Wootters and Zurek
1982).
Pitowsky took the uncertainty relations to be “the centerpiece that demarcates
between the classical and quantum domain” (2006, p. 214). The only non-classical
axiom in the Birkoff von Neumann axiomatization, and thus the logical anchor of the
uncertainty relations, is the axiom of irreducibility.21 While a classical probability
space is a Boolean algebra where for all events x and z (and where the event z⊥ is
the complement to z):
 
x = (x z) x z⊥ (reducibility)

in QM, we get irreducibility, i.e., (with 0 as the null event and 1 the certain event):
 
If for some z and for all x, x = (x z) x z⊥ then z = 0 or z = 1

Irreducibility signifies the non-Boolean nature of the algebra of possible events,

for the only irreducible Boolean algebra is the trivial one {0,1}. According to
Birkhoff and von Neumann, irreducibility means that there are no ‘neutral’ elements

20 The idea that the collapse can always be understood epistemically has been challenged by Meir
Hemmo; see Hagar and Hemmo (2006).
21 Pitowsky’s formulation is slightly different from that of Birkhoff and von Neumann, but the

difference is immaterial. Pitowsky makes significant progress, however, in his treatment of the
representation theorem for the axiom system, in particular in his discussion of Solér’s theorem.
The theorem, and the representation problem in general, is crucial for the application of Gleason’s
theorem (Gleason 1957), but will not concern us here.
5 Pitowsky’s Epistemic Interpretation of Quantum Mechanics and the PBR Theorem 111

z, z = 0 z = 1 such that for all x, x = (x z) U (x z⊥ ). If there would be such

‘neutral’ events, we would have non-trivial projection operators commuting with
all other projection operators. Intuitively, irreducibility embodies the uncertainty
relations for when x cannot be presented as the union of its intersection with z and
its intersection with z⊥ , then x and z cannot be assigned definite values at the same
time. Thus, whenever x = (x z) (x z⊥, )—x and z are incompatible and,
consequently, a measurement of one yields no information about the other. The
axiom further implies genuine uncertainty, or indeterminism—probabilities strictly
between (unequal to) 0 and 1. This result follows from a theorem Pitowsky calls
the logical indeterminacy principle, and which proves that for incompatible events
x and y

p(x) + p(y) < 2

The loss of information upon measurement—the phenomenon called ‘distur-

bance’ by the founders of QM—also emerges as a formal consequence of the
probabilistic picture. The interpretation of the uncertainty relations that emerges
from this axiomatization is obviously at variance with the ensemble interpretation
described in the previous section. Whereas the latter tolerates the ascription of well-
defined values to non-commuting operators/observables and construes the uncer-
tainty relations as constraining only the distribution of values over the ensemble,
the constraint derived from the Birkhoff- von Neumann-Pitowsky axiomatization
applies to the individual quantum system.
Having shown that his axiom system entails genuine uncertainty, Pitowsky
moved on to demonstrate the violation of the Bell-inequalities, i.e. the phenomena
of entanglement and nonlocality. Violations already appear in finite-dimensional
cases and follow from the calculation of the probabilities of the intersection of the
subspaces of the Hilbert space representing the (compatible) measurement results
at the two ends of the entangled system. Pitowsky showed, in both logical and
geometrical terms, that the quantum range of possibilities is indeed larger than
the classical range so that we get more correlation than is allowed by the classical
conditions. Whereas the standard response to this phenomenon consists in attempts
to discover the dynamic that makes it possible, Pitowsky emphasizes that this is a
logical-conceptual argument, independent of specific physical considerations over
and above those that follow from the non-Boolean character of the event structure.
He says
Altogether, in our approach there is no problem with locality and the analysis remains intact
no matter what the kinematic or the dynamic situation is; the violation of the inequality is
a purely probabilistic effect. The derivation of Clauser-Horne inequalities . . . is blocked
since it is based on the Boolean view of probabilities as weighted averages of truth values.
This, in turn, involves the metaphysical assumption that there is, simultaneously, a matter
of fact concerning the truth-values of incompatible propositions . . . . [F]rom our perspective
the commotion about locality can only come from one who sincerely believes that Boole’s
conditions are really conditions of possible experience. . . . .But if one accepts that one is
112 Y. Ben-Menahem

simply dealing with a different notion of probability, then all space-time considerations
become irrelevant (2006, pp. 231–232).

Recall that in order to countenance nonlocality without breaking with STR, the
no signaling constraint must be observed. As nonlocality is construed by Pitowsky
in formal terms—a manifestation of the quantum mechanical probability calculus,
uncommitted to a particular dynamic—it stands to reason that no signaling will
likewise be derived from probabilistic considerations. Indeed, it turns out that the no
signaling constraint can be understood as an instance of the more general principle
known as the non-contextuality of measurement (Barnum et al. 2000). In the spirit
of the probabilistic approach to QM, Pitowsky and Bub therefore maintain that
No signaling is not specifically a relativistic constraint on superluminal signaling. It is
simply a condition imposed on the marginal probabilities of events for separated systems
requiring that the marginal probability of a B-event is independent of the particular set of
mutually exclusive and collectively exhaustive events selected at A, and conversely” (Bub
and Pitowsky 2010, p. 443).22

To sum up: Pitowsky proposed a radical epistemic interpretation of QM. It is

epistemic in the sense that the probabilities represent partial knowledge (infor-
mation, rational belief) of the observer and it is radical in the sense that it is not
assumed to be knowledge (information, rational belief) about the physical state of
the system, but rather about observable measurement outcomes. It therefore deviates
from epistemic interpretations of probability in other areas such as gambling or
statistical mechanics, where the reality of underlying well-defined states is taken for
granted. Pitowsky justified the denial of the reality of the physical state by urging
that the assumption of determinate physical states leads to classical probabilities,
whereas an adequate systematization of quantum measurement results requires a
non-classical theory of probability. In the framework of this non-classical theory,
he claimed, the troubling features of QM—nonlocality in particular—are grounded
in the quantum formalism. No deeper explanation is needed. Just as the structure
of Minkowski’s spacetime accounts for relativistic effects that in earlier theories
received dynamical explanations, the structure of QM as a non-classical theory of
probability accounts for quantum effects that according to other interpretations call
for dynamical explanations.

22 Inthe literature, following in particular Jarrett 1984, it is customary to distinguish outcome

independence, violated in QM, from parameter independence, which is observed, a combination
that makes possible the peaceful coexistence with STR. The non-contextuality of measurement
amounts to parameter independence. See, however Redhead (1987) and Maudlin (2011), among
others, for a detailed exposition and critical discussion of the distinction between outcome and
parameter independence and its implications for the compatibility with STR. More recently,
Brandenburger and Yanofsky (2008) also refine Jarrett’s distinction. I thank Meir Hemmo for
drawing my attention to this paper. See also the paper by Levy and Hemmo in this volume.
5 Pitowsky’s Epistemic Interpretation of Quantum Mechanics and the PBR Theorem 113

5.4 The PBR Theorem

Pusey, Barrett and Rudolph summarize the no-go argument that has come to be
known as the PBR theorem in the following sentence:
We show that any model in which a quantum state represents mere information about
an underlying physical state of the system, and in which systems that are prepared
independently have independent physical states, must make predictions that contradict those
of quantum theory. (Pusey et al. 2012, p. 475)

Clearly, the theorem purports to target the epistemic interpretation of the wave
function, thereby supporting its realist interpretation. This is indeed the received
reading of the theorem and the source of its appeal. Under the title “Get Real” Scott
Aaronson (2012) announced the new result as follows:
Do quantum states offer a faithful representation of reality or merely encode the partial
knowledge of the experimenter? A new theorem illustrates how the latter can lead to a
contradiction with quantum mechanics. (2012, p. 443)

Aaronson distinguishes between the more common question of whether a quantum

state may correspond to different physical states, and the PBR question—whether
a physical state can correspond to different quantum states. It is the latter that the
PBR theorem purports to answer in the negative (See Fig. 5.1).
It is crucial to understand that we’re not discussing whether the same wavefunction can be
compatible with multiple states of reality, but a different and less familiar question: whether
the same state of reality can be compatible with multiple wavefunctions. Intuitively, the
reason we’re interested in this question is that the wavefunction seems more ‘real’ if the
answer is no, and more ‘statistical’ if the answer is yes (2012, p. 443).

Since the PBR theorem does indeed answer the said question in the negative, the
conclusion is that the wave function has gained (more) reality.
Why is the negative answer to this question a refutation of the epistemic
interpretation? Here Pusey et al. build on a distinction between Ψ -ontic models
and Ψ -epistemic models introduced by Harrigan and Spekkens (2010): In Ψ -

Quntum level

Physical level

(a) The same quantum state corresponds (b) The same physical state
to several physical states corresponds to two quantum states

Fig. 5.1 Two types of relation between the physical level and the quantum level. (a) The same
quantum state corresponds to several physical states. (b) The same physical state corresponds to
two quantum states
114 Y. Ben-Menahem

ontic models the Ψ function corresponds to the physical state of the system; in
Ψ -epistemic models, Ψ represents knowledge about the physical state of the system.
Consequently, there are also two varieties of incompleteness: Ψ could give us a
partial description of the physical state or a partial representation of our knowledge
about that state. If a Ψ -ontic model is incomplete, it is conceivable that Ψ could be
supplemented with further parameters—‘hidden variables’. In this case, the same Ψ
function could correspond to various physical states of the system, distinguishable
by means of the values of the additional hidden variables. Presumably, Ψ -epistemic
models can also be complete or incomplete but completing them cannot be accom-
plished by hidden variables of the former kind. Note that this analysis presupposes
a definite answer to the question posed at the end of the introduction to this
paper–what, according to the epistemic view, is the knowledge (belief, information)
represented by the quantum state about? The assumption underlying the theorem is
that it is knowledge about the physical state of the system.
So far, Harrigan and Spekkens’ distinction is merely terminological, but they
proceed to offer a criterion that distinguishes Ψ -epistemic from Ψ -ontic models:
If the Ψ function is understood epistemically, they claim, it can stand in a non-
functional relation to the physical state of the system, that is, the same physical state
may correspond to different (non-identical but also not orthogonal) Ψ functions.
Or, when probabilities rather than sharp values are considered, the supports of the
probability distributions corresponding to different Ψ functions can overlap for
some physical states (See Fig. 5.2 for an illustration of this criterion in the case
of two Ψ functions).23
Harrigan and Spekkens maintain that the possibility of overlap only makes
sense under the epistemic interpretation of Ψ , for, they argue, knowledge about
the physical state could be updated without any change to the physical state itself.

Δ
Ψ -ontic models Ψ -epistemic models

Fig. 5.2 The difference between Ψ -ontic and Ψ -epistemic models: (a) Ψ -ontic models. (b) Ψ -
epistemic models

23 Figure 5.2 follows the schematic illustration in the PBR paper.

5 Pitowsky’s Epistemic Interpretation of Quantum Mechanics and the PBR Theorem 115

In short, a criterion for a model being Ψ -epistemic is precisely the possibility of

such overlap of the supports of different probability functions. Getting ahead of
my argument, I should stress that allowing such a non-functional relation between
the physical state of the system and the quantum state, namely, allowing that the
quantum state is not uniquely determined by the physical state, is a non-trivial
assumption. The question of whether an epistemic theorist is actually committed
to it is crucial for assessing the verdict of the PBR theorem. In philosophical
terminology, the ontic option illustrated in figure 1(a), where every physical state
corresponds to a single quantum state is referred to as a relation of supervenience.
(There is no assumption of supervenience in the reverse direction). In figure 1(b),
however, illustrating what Harrigan and Spekkens call an epistemic model, the
quantum state does not even supervene on the physical state. This failure of
supervenience is not a necessary feature of probabilistic theories and is indeed
being debated in other physical theories involving probabilities, such as statistical
mechanics. I will return to this point below.
The authors of the PBR theorem make two assumptions: 1.The system has a
definite physical state—not necessarily identical with the quantum state. 2. Systems
that are prepared independently have independent physical states. PBR prove that
for distinct quantum states, if epistemic models are allowed, that is (using the
Harrigan Spekkens criterion), if an overlap of the probability distributions is allowed
so that the intersection of the supports has a non-zero measure—then a contradiction
with the predictions of QM could be derived. They derive the contradiction first for
√
the special case in which there are two systems such that | < Ψ 0 |Ψ 1 > | = 1/ 2 and
then generalize to multiple (non-orthogonal) states. In the special case the proof is
based on the following steps.
1. The authors consider two independent preparations of a system with states
|
0 > and |
1 > .
2. They choose a basis for the Hilbert space for which
√
|
0 > = |0 > and|
1 > = (|0 > +|1>)/ 2. (Such a basis can always be found).
3. They consider an epistemic model, namely, the probability distributions for the
two states overlap in a non-zero region Δ. Thus, there is a positive probability
q > 0 to get a physical state from that region in each one of the systems and,
because of independence, there is a positive probability q2 > 0 to get such a value
from the region Δ for both systems together.
4. They assume that the two systems are brought together and their entangled state
measured.
5. They show that QM assigns zero probability to all four orthogonal states onto
which such a measurements projects, whereas by the assumption made in step 3,
the probability is non-zero.
Contradiction!
The conclusion PBR draw is that epistemic models are incompatible with the
predictions of QM, and thus ruled out. We are therefore left with ontic models,
which are “more real,” to use Aaronson’s formulation.
116 Y. Ben-Menahem

5.5 Possible Responses by the Epistemic Theorist

Naturally, the strength of a theorem depends on the strength of its assumptions.

In the case of the PBR theorem, its second assumption—independence—has in
fact been called into question (Schlosshauer and Fine 2012, 2014), and will not be
discussed here. One could, however, also question the first assumption, namely, the
assumption that the system has a definite physical state and that in epistemic models
the quantum state represents the experimenter’s knowledge (belief, information)
about that physical state. Clearly, the assumption regarding the reality of a physical
state of the system is a realist assumption. The theorem therefore undermines a
particular combination of a realist assumption about physical states and a non-
realist interpretation of quantum states. The contradiction it derives may indicate
that this peculiar combination of realist and non-realist ingredients does not work.
From the logical point of view, one could also take the realist assumption itself to
be responsible for the contradiction, in which case the theorem could be taken to
challenge, rather than support a realist interpretation of quantum states. As we have
seen, the common reading of the theorem is that it demonstrates the failure of an
epistemic interpretation that allows a non-functional relation between the physical
state and the quantum state (i.e., it allows several quantum states corresponding
to a single physical state). It follows that an epistemic theorist who declines this
particular understanding of the epistemic approach, thereby rejecting the Harrigan
Spekkens criterion for epistemic models, will not be threatened by the theorem. Let
us consider these options in more detail.
Recall that the existence of definite physical states underlying quantum states
has been a matter of controversy ever since the early days of QM and yet,
Pusey et al. make no effort to motivate their assumption.24 They note briefly that
instrumentalists would deny it, implying that only instrumentalists would deny it
(and perhaps implying, further, that instrumentalism is too ridiculous a position
to be argued against). In his “Extended Review” of the literature on the reality of
the quantum state, Leifer (2014) elucidates the situation by distinguishing realist
epistemic interpretations (namely, those assigning well-defined physical states to
quantum systems) from “anti-realist, instrumentalist, or positivist” interpretations,
which he dubs “neo Copenhagen” (2014, p. 72). He acknowledges that neo-
Copenhagen interpretations are not affected by the PBR theorem, but, like the
authors of the theorem, he is somewhat dismissive about them and more importantly,
he does not even consider the possibility of epistemic interpretations at variance with
instrumentalism. We will see, however (Sect. 5.6), that Pitowsky’s epistemic inter-
pretation, although denying the definite-state assumption, still differs significantly
from instrumentalism. Moreover, Leifer construes epistemic states as subjective
in the sense that an epistemic state is “something that exists in the mind of the

24 Interpretationssuch as Bohm’s theory and the many world interpretation show that the
assumption is not ruled out by QM, but it is, nonetheless, sufficiently contentious to require some
justification.
5 Pitowsky’s Epistemic Interpretation of Quantum Mechanics and the PBR Theorem 117

observer rather than in the external physical world,” (2014, p. 69) and can therefore
vary from one observer to another. Pitowsky, I suggest, did not understand the
epistemic interpretation in that way. QM, he thought, is an algorithm that determines
degrees of rational belief, or rational betting procedures in the quantum context.
Even if the belief is in an individual’s mind, the algorithm is objective. According
to an epistemic interpretation of this kind, an observer using classical probability
theory rather than QM, is bound to get results that deviate from QM, or, in the
betting jargon often adopted by epistemic theorists, is bound to lose. The epistemic
interpretation construes quantum probabilities as constraints on what can be known,
or rationally believed, about particular measurement outcomes. By itself, it does not
imply more liberty (about the ascription of probabilities to certain outcomes) than
realist interpretations.
To what extent is an epistemic theorist committed to the Harrigan-Spekkens con-
dition of non-supervenience as characterizing Ψ -epistemic models? Let’s assume
that the condition holds and there are well-determined physical states that ‘corre-
spond’ to different quantum states. Obviously, this correspondence is not meant to
be an identity for a physical state cannot be identical with different quantum states.
A possible way of unpacking this correspondence is as follows: a well-defined
physical state exists (prior to measurement) and, upon measurement ‘generates’
or ‘gives rise to’ different results. Such as interpretation is at odds with the
received understanding of Bell’s inequalities and, more specifically, with Pitowsky’s
interpretation of QM as a non-classical theory of probability. We have seen that
Pitowsky explicitly denied the classical picture of well-defined physical states
generating quantum states, a picture that leads to Bool’s classical conditions, which,
in turn, are violated by QM. Indeed, the language of a physical state ‘giving rise’
to, or ‘generating,’ the result of a quantum measurement would be rejected by
Pitowsky even if it referred to a physical state corresponding to single quantum
state, not merely to different ones as is the assumption in the Harrigan-Spekkens
analysis. If, on the other hand, by the said ‘correspondence’ we mean the physical
state of the system at the very instant it is being measured (or thereafter, but before
another measurement is made), we are no longer running into conflict with Bell’s
theorem, but then, there is no reason to accept that this physical state corresponds to
different quantum states. In other words, there is no reason to deny supervenience.
The Epistemic theorist can take quantum states to determine the probabilities of
measurement results, and understand probabilities as degrees of rational belief,
and still deny the possibility that the same physical state corresponds to different
quantum states. It would actually be very strange if the epistemic theorist would
endorse such a possibility. Clearly, the quantum state—a probability amplitude—is
compatible with different measurement outcomes, but the probability itself is fixed
by QM.
Moreover, if quantum probabilities are conceived as the maximal information
QM makes available, then surely we cannot expect different probability amplitudes
118 Y. Ben-Menahem

to correspond the same physical state.25 The idea of maximal information plays
an important role in Spekkens (2005), where, on the basis of purely information-
theoretic constraints, a toy model of QM is constructed. The basic principle of this
toy model is that in a state of maximal knowledge, the amount of knowledge is equal
to the amount of uncertainty, that is, the number of questions about the physical state
of a system that can be answered is equal to the number of questions that receive
no answer. (This was also Schrödinger’s intuition cited in the previous footnote).
Spekkens succeeds in deriving from this principle many of the characteristic
features of QM, for example, interference, disturbance, the noncommutative nature
of measurement and no cloning.26 Either way, then—whether the real physical
state allegedly corresponding to different quantum states is taken to exist prior to
measurement or thereafter—the epistemic theorist I have in mind—Pitowsky in
particular—need not accept the Harrigan-Spekkens condition, let alone endorse it
as a criterion for epistemic models.
A comparison with statistical mechanics is again useful. The goal of statistical
mechanics is to derive thermodynamic behavior, described in terms of macro-
properties such as temperature, pressure and entropy, from the description of
the underlying level of micro-states in terms of classical mechanics. For some
macroscopic parameters, the connection with micro-properties is relatively clear—
it is quite intuitive, for instance, to correlate the pressure exerted by a gas on
its container with the average impact (per unit area) of micro-particles on the
container. But this is not the case of other macro-properties; entropy, in particular.
Recovering the notion of entropy is essential for the recovery of the second law
of thermodynamics and constitutes a major challenge of statistical mechanics. The
fundamental insight underlying the connection between entropy and the micro-

25 The idea of maximal information or completeness is explicit in Schrödinger, who took the
Ψ function to represent a maximal catalog of possible measurements. The completeness of the
catalog, he maintained, entails that we cannot have a more complete catalog, that is, we cannot
have two Ψ functions of the same system one of which is included in the other. For Schrödinger
this is the basis for ‘disturbance’: Any additional information, arrived at by measurement, must
change the previous catalog by deleting information from it, meaning that at least some of the
previous values have been destroyed. Entanglement is also explained by Schrödinger on the basis
of the maximality or completeness of
. He argues as follows: A complete catalog for two separate
systems is, ipso facto, also a complete catalog of the combined system, but the reverse does not
follow. “Maximal knowledge of a total system does not necessarily include total knowledge of
all its parts, not even when these are fully separated from each other and at the moment are not
influencing each other at all” (1935, p. 160, italics in original). The reason we cannot infer such
total information is that the maximal catalog of the combined system may contain conditional
statements of the form: if a measurement on the first system yields the value x, a measurement on
the second will yield the value y, and so on. Schrödinger sums up: “Best possible knowledge of
a whole does not necessarily include the same for its parts . . . The whole is in a definite state, the
parts taken individually are not” (1935, p. 161). In other words, separated systems can be correlated
or entangled via the Ψ function of the combined system, but this does not mean that their individual
states are already determined.
26 Spekkens also notes that the model fails to reproduce the Kochen-Specker theorem and the

violation of Bell-inequalities. In that sense it is still a toy model.

5 Pitowsky’s Epistemic Interpretation of Quantum Mechanics and the PBR Theorem 119

level was that macrostates are multiply realizable by microstates. The implication is
that in general, since the same macrostate could be realized by numerous different
microstates, the detailed description of the microstate of the system is neither
sufficient nor necessary for understanding the properties of its macrostate. What
is crucial for the characterization of a macrostate in terms of the micro-structure
of the system realizing it, however, is the number of ways (or its measure-theoretic
analogue for continuous variables) in which a macrostate can be realized by the
microstates comprising it. Each macrostate is thus realizable by all the microstates
corresponding to points that belong to the phase-space volume representing this
macrostate—a volume that can vary enormously from one macrostate to another.
Note the nature of the correspondence between microstates and macrostates in this
case: When considering a particular system, at each moment it is both in a specific
microstate and in a specific macrostate. In other words, microstates neither give rise
to macrostates nor generate them! The system is in single well-defined state, but it
can be described either in terms of its micro-properties or in terms of its macro-
properties. For an individual state, the distinction between macro and micro levels
is not a distinction between different states but rather between different descriptions
of the same state. It is only the type, characterized in terms of macro-properties, that
has numerous different microstates belonging to it.27
The insight that macrostates differ in the number of their possible realizations (or
it measure-theoretic analogue) led to the identification of the volume representing
a macrostate in phase space with the probability of this macrostate and to the
definition of entropy in terms of this probability. On this conception, the maximal
entropy of the equilibrium state in thermodynamic terms is a manifestation of
its high probability in statistical-mechanical terms.28 The link between entropy
and probability constituted a crucial step in transforming thermodynamics into
statistical mechanics, but it was not the end of the story. For one thing, the
meaning of probability in this context still needed refinement, for another, the
connection between the defined probability and the dynamic of the system had to be
established.29

27 Itseems that this is not the way Harrigan and Spekkens or Puesy et al. conceive of the relation
between the physical state and the quantum state. This obscurity might be the source of the
problems raised here regarding their characterization of the epistemic interpretation.
28 The probability of a macrostate can be understood objectively in terms of the ‘size’ of that

macrostate (in phase space) or subjectively, as representing our ignorance about the actual
microstate (as in the above quotation from Jaynes). This difference in interpretation does not
affect the probabilities of macrostates. Note that from the perspective of a single microstate (and
certainly from the perspective of a single atom or molecule), it should not matter whether it belongs
to a macrostate of high entropy, that is, a highly probable macrostate to which numerous other
microstates also belong, or to a low-entropy state—an improbable one, to which only a tiny fraction
of microstates belongs. In so far as only the microlevel is concerned, entropy does not constitute
an observable physical category.
29 A natural account of probability would involve an ensemble of identical systems (possibly an

infinite ensemble), where probabilities of properties are correlated with fractions of the ensemble.
In such an ensemble of systems we should expect to find more systems in probable macrostates
120 Y. Ben-Menahem

Different approaches to statistical mechanics differ in their response to these

problems. Whereas Boltzmann’s approach is more focused on the dynamical
evolution of the individual system, in Gibbs’ statistical mechanics it is the math-
ematical precision of the notion of probability, and thus the ensemble of identical
systems, that takes priority. One of the differences between these approaches is that
in Boltzmann’s approach macrostates supervene on microstates, while in Gibbs’
ensemble approach they do not. The epistemic interpretation of quantum states as
construed in the PBR theorem is in this respect closer to Gibbs’ statistical mechanics
than to Boltzmann’s. But since the PBR theorem rules out this kind of epistemic
interpretation, it ultimately points to a disanalogy between OM and Gibbs’ statistical
mechanics. At the same time, the epistemic interpretation (as understood by the
PBR theorem), also differs from Boltzmann’s statistical mechanics in allowing the
failure of supervenience in the first place, and exposes once again a disanalogy
between QM and statistical mechanics. By contrast, what Harrigan and Spekkens
call an ontic model does supervene on the physical state of the system and is in that
sense closer to Boltzmann’s statistical mechanics than to Gibbs’. But we are then
left with the above mentioned difficulties concerning the legitimacy of postulating
a level of well-defined physical states underlying quantum states in the same way
that microstates in statistical mechanics underlie macrostates. Whichever approach
to statistical mechanics we choose, the analogy between statistical mechanics and
QM does not seem to work.30
These considerations are sufficient to indicate that the PBR account of the
epistemic interpretation as taking quantum states to represent knowledge (belief,
information) about the physical state of the system—especially when combined
with the Harrigan -Spekkens condition of non supervenience—is not as innocent
as it appears. Pitowsky’s interpretation exemplifies an epistemic view that does
not share the PBR assumption of non-supervenience and is therefore not refuted
by the theorem. It could now be objected that there was no need to go to such
length to defend Pitowsky’s view: the restriction to the surface level of measurement
outcomes amounts to instrumentalism, a position that according to Pusey et al. is
immune to their theorem. Let me now turn to this objection.

than in improbable ones. And yet, how does this consideration reflect on the individual system
we are looking at? Assuming an actual system that happens to be in an improbable macrostate—it
would certainly not follow from the above probabilistic considerations that it should, in the course
of time, move from its improbable state to a more probable one. If one’s goal is to account for the
evolution of an individual system (in particular, its evolution to equilibrium), there is a gap in the
argument. In addition, there is the formidable problem of deriving thermodynamic irreversibility
from the statistical-mechanical description, which is based on the time-reversal-symmetric laws of
mechanics. Since this problem has no analogue in the present discussion of the interpretation of
quantum states, we can set it aside.
30 As mentioned in Sect. 5.1, the analogy was rejected by the Copenhagen school from its earliest

days.
5 Pitowsky’s Epistemic Interpretation of Quantum Mechanics and the PBR Theorem 121

5.6 Instrumentalism

Admittedly, the instrumentalist will not assume the existence of physical states that
cannot be observed and will thus avoid the first assumption of the PBR theorem.
In this respect instrumentalism and Pitowsky’s radical epistemic interpretation
are alike. But here they part ways. To see why, it is useful to characterize
instrumentalism more precisely as committed to a verificationist (non-realist) theory
of meaning.31 On this theory, the meaning of a sentence depends on the existence
of a verification procedure—where none exists, the sentence is meaningless—and
its truth (falsity) on the existence of a proof (refutation). Unlike the realist, who
accepts bivalence (every sentence is either true or false regardless of whether or
not we can prove it), the verificationist admits a third category of indeterminate
sentences—sentences that are neither provable, nor disprovable. Consider the
difference between realism and verificationism when entertaining the claim that
Julius Cesar, say, could not stand raw fish or that he had a scar on his left toe.
Assuming that there is no evidence to settle the matter, the verificationist will deem
these claims indeterminate, whereas the realist will insist that although we may
never know, they must be either true or false. What if we stipulate that Julius
Cesar detested raw fish, or that he did in fact have a scar on his toe? On the
assumption that there is no evidence, such stipulation cannot get us into conflict
with the facts. We are therefore free to stipulate the truth values of indeterminate
sentences. This freedom, I claim, is constitutive of indeterminacy. The continuum
hypothesis provides a nice example from mathematics: Since it is independent of
the axioms of set theory, it might be true in some models of set theory and false
in others. If by stipulating the truth value of the hypothesis we could generate a
contradiction with set theory, we could no longer regard it as undecidable. The
axiom of parallels provides another example of the legitimacy of stipulation. Since
it is independent of the other axioms its negation(s) can be added to the other axioms
without generating a contradiction. In QM, questions regarding the meaning of
quantum states were for a long time undecided. If they were indeed undecidable,
that is undecidable in principle, the verificationist (instrumentalist) could stipulate
the answer by fiat. Schrödinger provided the first clue that this is not the case—the
very stipulation of determinate values would lead to conflict with QM. Later no-go
theorems and experiments confirmed this conclusion. As we saw, Pitowsky argued
that the reason for the violation of Bell’s inequalities is that in order to derive the
inequalities, one makes the classical assumption of determinate states underlying
quantum states. The violation of the inequalities, he thought, indicates that this
classical assumption is not merely indeterminate but actually false—leading to a
classical rather than quantum probability space. The constitutive characteristic of

31 In this section I repeat an argument made in Ben-Menahem (1997). The difference between
instrumentalism and the radical epistemic interpretation stands even for an instrumentalist who
does not explicitly endorse verificationism, for even such an instrumentalist would not envisage
the empirical implications that the radical epistemist calls attention to.
122 Y. Ben-Menahem

freedom is therefore missing: If the stipulation of definite states leads to conflict

with QM, it is no longer up to the instrumentalist to approve of their existence or
deny it. The radical epistemic interpretation proposed by Pitowsky is not neutral
with regard to the assumption of definite physical states. Unlike the instrumentalist,
who assumes there is no fact of the matter, and thus withholds judgement, radical
epistemic theorists take a definite position, albeit a negative one. They deny the
assumption and expect this denial to have empirical implications. To identify the
negative with the neutral is a category mistake: There is a fundamental difference
between provable negative assertions—no largest prime number—and no-fact-of-
the-matter assertions.
I have reviewed four accounts of quantum probabilities: The Copenhagen
conception—a paradoxical ‘Verschmelzung’ of probabilistic and physical elements;
the ensemble interpretation, which is probabilistic throughout; the epistemic con-
ception assumed by PBR; the radical epistemic interpretation proposed by Pitowsky.
These accounts correspond to the following responses to the question of what quan-
tum probabilities refer to (respectively): observable states of individual systems,
distribution of physical states in an ensemble of similar systems, the observer’s
knowledge about the physical state of an individual system, objective constraints on
the information that can be gleaned from QM on possible measurement outcomes.
The ‘knowledge about what?’ question clearly distinguishes the epistemic inter-
pretation discussed by PBR from the radical interpretation proposed by Pitowsky.
Several differences between the role of probabilities in QM and statistical mechanics
have been noted, primarily differences regarding the supervenience of higher-level
states on an underlying level of well-defined physical states. We have seen that,
due to their impact on the meaning of fundamental quantum laws (the uncertainty
relations in particular), questions regarding the meaning of probability are vital in
the context of QM, but not in statistical mechanics, where different interpretations of
probability do not change the contents of the theory. Finally, the difference between
instrumentalism and Pitowsky’s radical epistemic interpretation has been clarified.
The PBR theorem targets an epistemic interpretation that combines a realist
assumption—the existence of definite physical states underlying quantum states—
with a non-realist construal of quantum probabilities. Inspired by Pitowsky, I have
argued that this is not the only available epistemic interpretation, and not the one
most commonly upheld by epistemic theorists. An interpretation polar to that of
PBR is available: One can let go of the assumption of definite physical states,
but construe quantum probabilities as objective constraints on the information
made available by measurement. From the perspective of such a radical epistemic
interpretation, the Harrigan-Spekkens criterion of non-supervenience does not make
sense. If I am right, the widespread reading of the PBR theorem as dealing a fatal
blow to the epistemic interpretation should be reconsidered.
5 Pitowsky’s Epistemic Interpretation of Quantum Mechanics and the PBR Theorem 123

References

Aaronson, S. (2012). Get real. Nature Physics, 8, 443–444.

Aharonov, Y., & Bohm, D. (1959). Significance of electromagnetic potentials in the quantum
theory. Physics Review, 115, 485–489.
Aharonov, Y., & Rohrlich, D. (2005). Quantum paradoxes: Quantum theory for the perplexed.
Winheim: Wiley-VCH.
Albert, D. (2000). Time and chance. Cambridge, MA: Harvard University Press.
Ballentine, L. E. (1970). The statistical interpretation of quantum mechanics. Review of Modern
Physics, 42, 358–381.
Barnum, H., Caves, C. M., Finkelstein, J., Fuchs, C. A., & Schack, R. (2000). Quantum probability
from decision theory? Proceedings of the Royal Society of London A, 456, 1175–1190.
Ben-Menahem, Y. (1997). Dummett vs Bell on quantum mechanics. Studies in History and
Philosophy of Modern Physics, 28, 277–290.
Ben-Menahem, Y. (2012). Locality and determinism: The odd couple. In Y. Ben-Menahem &
M. Hemmo (Eds.), Probability in physics (Frontiers in Science Series) (pp. 149–166). Berlin:
Springer.
Ben-Menahem, Y. (2017). The PBR theorem: Whose side is it on? Studies in History Philosophy
of Modern Physics, 57, 80–88.
Birkhoff, G., & von Neumann, J. (1936). The logic of quantum mechanics. Annals of Mathematics,
37, 823–845.
Blokhintsev, D. I. (1968). The philosophy of quantum mechanics. Amsterdam: Reidel.
Born, M. [1926] (1963). Quantenmechanik der Stossvorgange. Zeitschrift f. Physik, 38, 803–827.
In: Ausgewahlte Abhandlungen II (233–252). Gottingen: Vandenhoek Ruprecht.
Brandenburger, A., & Yanofsky, N. (2008). A classification of hidden variable properties. Journal
of Physics A. Mathematical and Theoretical, 41(42)5302, 1–21.
Bub, J., & Pitowsky, I. (2010). Two dogmas about quantum mechanics. In S. Saunders, J. Barrett,
A. Kent, & D. Wallace (Eds.), Many worlds? Everett, quantum theory and reality (pp. 433–
459). Oxford: Oxford University Press.
De Finetti, B. (1974). Theory of probability. New York: Wiley.
Einstein, A., Podolsky, B., & Rosen, N. (1935). Can quantum-mechanical description of physical
reality be considered complete? Physics Review, 47, 777–780.
Feynman, R. P. (1951). The concept of probability in quantum mechanics. In Second Berkeley
symposium on mathematical statistics and probability, pp. 533–541.
Gleason, A. M. (1957). Measurement on closed subspaces of a Hilbert space. Journal of
Mathematics and Mechanics, 6, 885–893.
Hagar, A., & Hemmo, M. (2006). Explaining the unobserved—Why quantum mechanics ain’t only
about information. Foundations of Physics, 36, 1295–1324.
Harrigan, N., & Spekkens, R. W. (2010). Einstein, incompleteness, and the epistemic view of
quantum states. Foundations of Physics, 40, 125–157.
Healey, R. (2007). Gauging what’s real. Oxford: Oxford University Press.
Hemmo, M., & Shenker, O. (2012). The road to Maxwell’s demon. Cambridge: Cambridge
University Press.
Hetzroni, G. (2019a). Gauge and ghosts (forthcoming). The British Journal for the Philosophy of
Science. https://doi.org/10.1093/bjps/axz021.
Hetzroni, G. (2019b). The quantum phase and quantum reality. Ph.D. dissertation (submitted to
the Hebrew University of Jerusalem.
Kochen, S., & Specker, E. P. (1967). The problem of hidden variables in quantum mechanics.
Journal of Mathematics and Mechanics, 17, 59–89.
Leifer, M. S. (2014). Is the quantum state real? An extended review of Ψ –ontology theorems.
Quanta, 3, 67–155.
Maudlin, T. (2011). Quantum non-locality and relativity. Oxford: Wiley-Blackwell.
Park, J. (1970). The concept of transition in quantum mechanics. Foundations of Physics, 1, 23–33.
124 Y. Ben-Menahem

Pitowsky, I. (1989). Quantum probability, quantum logic (Lecture notes in physics) (Vol. 321).
Heidelberg: Springer.
Pitowsky, I. (2006). Quantum mechanics as a theory of probability. In Demopoulos & Pitowsky
(Eds.), Physical theory and its interpretation (pp. 213–239). Dordrecht: Springer.
Popper, K. R. (1967). Quantum mechanics without ‘The observer’. In M. Bunge (Ed.), Quantum
theory and reality (pp. 7–43). Berlin: Springer.
Popper, K. R. (1968). The logic of scientific discovery (Revised edn). London: Hutchinson.
Pusey, M. F., Barrett, J., & Rudolph, T. (2012). On the reality of the quantum state. Nature Physics,
8, 475–478.
Ramsey, F. P. (1931). Truth and probability. In R. B. Braithwait (Ed.), The foundations of
mathematics and other logical essays (pp. 156–198). London: Kegan Paul.
Redhead, M. (1987). Incompleteness, nonlocality and realism. Oxford: Clarendon Press.
Savage, L. J. (1954). The foundations of statistics. New York: Wiley.
Schlosshauer, M., & Fine, A. (2012). Implications of the Pusey-Barrett-Rudolph quantum no-go
theorem. Physical Review Letters, 108, 260404.
Schlosshauer, M., & Fine, A. (2014). No-go theorem for the composition of quantum systems.
Physical Review Letters, 112, 070407.
Schrödinger, E. (1935). The present situation in quantum mechanics (trans. J. D. Trimmer).
In: Wheeler & Zurek (Eds.), Quantum theory and measurement (pp. 152–167). Princeton:
Princeton University Press.
Spekkens, R. W. (2005). In defense of the epistemic view of quantum states: A toy theory.
arXive:quant-ph/0401052v2.
Wootters, W., & Zurek, W. H. (1982). A single quantum cannot be cloned. Nature, 299, 802–803.
Chapter 6
On the Mathematical Constitution
and Explanation of Physical Facts

Joseph Berkovitz

Abstract The mathematical nature of modern physics suggests that mathematics is

bound to play some role in explaining physical reality. Yet, there is an ongoing
controversy about the prospects of mathematical explanations of physical facts
and their nature. A common view has it that mathematics provides a rich and
indispensable language for representing physical reality but that, ontologically,
physical facts are not mathematical and, accordingly, mathematical facts cannot
really explain physical facts. In what follows, I challenge this common view. I argue
that, in modern natural science, mathematics is constitutive of the physical. Given
the mathematical constitution of the physical, I propose an account of explanation in
which mathematical frameworks, structures, and facts explain physical facts. In this
account, mathematical explanations of physical facts are either species of physical
explanations of physical facts in which the mathematical constitution of some
physical facts in the explanans are highlighted, or simply explanations in which
the mathematical constitution of physical facts are highlighted. In highlighting the
mathematical constitution of physical facts, mathematical explanations of physical
facts make the explained facts intelligible or deepen and expand the scope of
our understanding of them. I argue that, unlike other accounts of mathematical
explanations of physical facts, the proposed account is not subject to the objection
that mathematics only represents the physical facts that actually do the explanation.
I conclude by briefly considering the implications that the mathematical constitution
of the physical has for the question of the unreasonable effectiveness of the use of
mathematics in physics.

Keywords Mathematical explanations · Mathematical constitution of the

physical · Relationships between mathematics and physics · Effectiveness of

J. Berkovitz ()
Institute for the History and Philosophy of Science and Technology, University of Toronto,
Victoria College, Toronto, ON, Canada
e-mail: joseph.berkovitz@utoronto.ca

© Springer Nature Switzerland AG 2020 125

M. Hemmo, O. Shenker (eds.), Quantum, Probability, Logic,
Jerusalem Studies in Philosophy and History of Science,
https://doi.org/10.1007/978-3-030-34316-3_6
126 J. Berkovitz

mathematics in physics · Mathematical representations · Pythagorean ·

neo-Kantian

6.1 The Orthodoxy

Modern physics is highly mathematical, and its enormous success seems to suggest
that mathematics is very effective as a tool for representing the physical realm.
Nobel Laureate Eugene Wigner famously expressed this view in “The unreasonable
effectiveness of mathematics in the natural sciences”:
the mathematical formulation of the physicist’s often crude experience leads in an uncanny
number of cases to an amazingly accurate description of a large class of phenomena (Wigner
1960, p. 8).

The great success of mathematical physics has led many to wonder about the causes
of and reasons for the effectiveness of mathematics in representing the physical
realm. Wigner thought that this success is unreasonable and even miraculous.
The miracle of the appropriateness of the language of mathematics for the formulation of
the laws of physics is a wonderful gift which we neither understand nor deserve. (Ibid.,
1960, p. 11)

Mathematics is commonly conceived as the study of purely abstract concepts

and structures and the question is how such concepts and structures could be so
successful in representing the physical realm.1
The ubiquity and great success of mathematics in physics does not only raise
the puzzle of the “unreasonable” effectiveness of mathematics in physics. It also
suggests that mathematics is bound to play a role in explaining physical reality. Yet,

1 Steiner (1998, p. 9) argues that there are two problems with Wigner’s formulation of question of
the “unreasonable effectiveness of mathematics in the natural sciences.” First, Wigner ignores the
cases in which “scientists fail to find appropriate mathematical descriptions of natural phenomena”
and the “mathematical concepts that never have found an application.” Second, Wigner focuses
on individual cases of successful applications of mathematical concepts, and these successes
might have nothing to do with the fact that a mathematical concept was applied. Steiner seeks
to formulate the question of the astonishing success of the applicability of mathematics in the
natural sciences in a way that escapes these objections. He argues that in their discoveries of new
theories, scientists relied on mathematical analogies. Often these analogies were ‘Pythagorean’,
meaning that they were inexpressible in any other language but that of pure mathematics. That
is, often the strategy that physicists pursued to guess the laws of nature was Pythagorean: “they
used the relations between the structures and even the notations of mathematics to frame analogies
and guess according to those analogies” (ibid., pp. 4–5). Steiner argues that although not every
guess, or even a large percentage of the guesses, was correct, this global strategy was astonishingly
successful.
Steiner’s reasoning and examples are intriguing and deserve an in-depth study, which, for want
of space, I need to postpone for another opportunity. I believe, though, that such a study will not
change the main thrust of my analysis of the question of the mathematical constitution of the
physical and its implications for the questions of mathematical explanations of physical facts and
the “unreasonable effectiveness of mathematics in the natural sciences”.
6 On the Mathematical Constitution and Explanation of Physical Facts 127

there is an ongoing controversy about the prospects and nature of mathematical

explanations of physical facts. (Henceforth, the term ‘physical fact’ is meant to
subsume all aspects of the physical, such as laws, principles, properties, relations,
etc. Further, for the sake of simplicity and continuity with the literature, I will
use the term ‘physical facts’ to be more general so as to include natural facts.) A
common view has it that mathematics provides a rich and indispensable language
for representing physical reality but could not play a role in explaining physical
facts. A related prevalent view is that, ontologically, the physical is to be sharply
distinguished from the mathematical. Thus, it is common to think that sharing
mathematical properties does not entail sharing physical properties. The idea is
that, fundamentally, physical facts are not mathematical, and that mathematics only
provides a language for representing the physical realm, even if this language is
indispensable in the sense of being by far the most effective. These common views
about the nature of physical facts and the role that mathematics plays in physics
seem to be dogmas of contemporary mainstream schools of philosophy of science.
Accordingly, the idea that mathematical facts could explain physical facts seems
perplexing: How could facts about abstract, non-physical entities, which obtain in
all ‘possible worlds’ (including ‘non-physical’ ones) and are ontologically separated
from the physical, explain physical facts?
It should not be surprising then that until recently the subject of mathematical
explanations of physical facts drew little attention in the philosophical literature.
Yet, a recent interest in philosophical questions concerning the application of
mathematics in the natural sciences has led philosophers to study the prospects
and nature of mathematical explanations of physical facts, and various accounts
of such explanations have been proposed (see, for example, Steiner 1978; Clifton
1998; Colyvan 2001, 2002; Batterman 2002, 2010, 2018; Pincock 2004, 2011a, b,
2012, 2015; Baker 2005, 2009, 2017; Bokulich 2008a, b, 2011; Dorato and Felline
2011; Lyon 2012; Batterman and Rice 2014; Lange 2016, 2018; Baron et al. 2017;
Bueno and French 2018; Felline 2018; and Mancosu 2018). These accounts are
clearly intended to offer non-causal explanations of physical facts. Yet, the exact
nature of the explanations on offer is unclear. For example, some of these accounts
attempt to show that mathematical facts or mathematical structures ‘could make
a difference’ to physical facts (Bokulich 2008a, b, 2011; Pincock 2011b, 2015;
Baron et al. 2017). But granted the common view that, ontologically, there is a
sharp distinction between the physical and the mathematical, it is not clear how
mathematical structures or mathematical facts could make such a difference and
what the nature of the difference making is.
Further, the term ‘mathematical explanation of physical facts’ is ambiguous. In
what follows, by this term I mean explanations in which mathematical frameworks,
structures, or facts explain physical facts, and by ‘mathematical facts’ I mean
mathematical truths, such as 2 + 2 = 4. This kind of explanation is to be
distinguished from explanations of physical facts that merely appeal to mathematics
in order to represent physical facts. In the literature, there are various putative
examples of mathematical explanations of physical facts and attempts to subsume
them under a general account. In Sect. 9, I consider four such accounts. I argue that
128 J. Berkovitz

the nature of explanation in these accounts is unclear and that they are subject to the
objection that the mathematical frameworks, structures, or facts they appeal to play
a representational rather than explanatory role.

6.2 An Alternative Perspective

Although those who advocate the existence of mathematical explanations of phys-

ical facts endorse the idea of the indispensability of mathematics in physics, they
seem to presuppose that mathematics does not play a constitutive role in defining
what the physical is. I shall propose below that a way to circumvent the conundrum
that current accounts of mathematical explanations of physical facts encounter is to
acknowledge that, in modern natural science, the mathematical is constitutive of the
physical. I believe that there are good reasons to think that contemporary natural
science is in effect committed to such constitution. In particular, I will argue that
if physical facts are fundamentally non-mathematical, the common wisdom about
how mathematical models represent physical systems – which is based on the so-
called ‘mapping account’ – becomes a mystery. Accordingly, the idea that there is a
mathematical constitution of the physical should appeal to anybody who accepts the
mapping account as a necessary part of mathematical representation of physical
facts, independently of their view about whether mathematical explanations of
physical facts exist. In fact, reflecting on the nature of the physical in contemporary
natural science, I will argue that this idea should also be compelling to those who
reject the mapping account as a necessary part of mathematical representations.
Given the mathematical constitution of the physical, I shall propose a new
account of mathematical explanation of physical facts. The main idea of this account
is that a mathematical explanation of physical facts highlights the mathematical
constitution of some physical facts, thus making the explained facts intelligible
or deepening and expanding the scope of our understanding of them. While this
account could be applied to many of the putative examples of mathematical
explanations of physical facts in the literature, the question whether it could account
for all kinds of putative mathematical explanations of physical facts is beyond
the scope of this paper for various reasons. First, as mentioned above, the term
‘mathematical explanation of physical facts’ is ambiguous and my understanding of
it need not agree with all of its applications in the literature. Second, for lack of space
and to keep things as simple as possible, I will only sketch the proposed account and
consider it mainly in the context of simple examples of mathematical explanations
of physical facts. Third, the range of putative examples I have encountered is not
sufficiently broad to support such a grand claim. Yet, I think that there are good
reasons to believe that the proposed account could be applied to all the examples
of mathematical explanations of physical facts in which mathematical frameworks,
structures, or facts are supposed to explain physical facts.
6 On the Mathematical Constitution and Explanation of Physical Facts 129

The current study was inspired by an exchange that I had with participants of
the workshop The Role of Mathematics in Science at the Institute for the History
and Philosophy of Science and Technology at the University of Toronto in October
2010. To my surprise, the common view was that while mathematics is very
important for representing physical systems/states/properties/laws, it does not play
any constitutive role in physics. Even more surprising, this view was also shared
by supporters of ontic structural realism, for whom such a constitution seems to
be a natural premise. The workshop Mathematical and Geometrical Explanations at
the Universitat Autònoma de Barcelona in March 2012 and subsequent conferences
and colloquia provided me with the opportunity to present the proposed account of
mathematical explanation of physical facts, and the invitation to contribute to the
second volume in honor of Itamar Pitowsky prompted me to write up this paper.
Pitowsky had a strong interest in mathematics and its use in physics. His research
covers various topics in these fields: the philosophical foundations of probability;
the nature of quantum probability and its relation to classical probability and quan-
tum logic; the interpretation of quantum mechanics, in general, and as a theory of
probability, in particular; quantum non-locality; computation complexity and quan-
tum computation; and probability in statistical mechanics. This research has inspired
a new perspective on the non-classical nature of quantum probability, quantum logic,
and quantum non-locality, and influenced our understanding of these issues.2
Pitowsky’s view of the relationship between mathematics and physics is less
known, however. Based on his commentary on Lévy-Leblond’s reflections on the
use and effectiveness of mathematics in physics in the Bar-Hillel colloquium in the
mid 1980s, in the next section I present his thoughts on this topic. Lévy-Leblond’s
and Pitowsky’s reflections will provide a background for the main focus of the
paper: the mathematical constitution and mathematical explanations of physical
facts. Both Lévy-Leblond and Pitowsky embrace the idea of the mathematical
constitution of the physical. Yet, it is not clear from their discussion what the exact
nature of this constitution is. In Sect. 4, I review two traditional conceptions of
mathematical constitution of physical facts – the Pythagorean and the neo-Kantian –
and briefly give an example of mathematical constitution of physical facts. In
Sect. 5, I review the common view of how mathematical models/theories represent
physical facts. The main idea of this view is that a mathematical model/theory
represents the properties of its target physical entities on the basis of a similarity
between the relevant mathematical structure in the model/theory and the structure of
the properties of these entities. More precisely, the idea is that there is an appropriate
structure-preserving mapping between the mathematical system in the model/theory
and the properties and relations of the represented physical entities. I argue that
the mapping account undermines the idea that physical facts are non-mathematical.
Further, in Sect. 6, I argue that a reflection on the concept of the physical in modern
natural science supports the idea that the physical is constituted by the mathematical.
In Sect. 7, I suggest that the philosophical interpretative frameworks of scientific

2 For a list of Pitowsky’s publications, see https://www.researchgate.net/scientific-contributions/

72394418_Itamar_Pitowsky
130 J. Berkovitz

theories/models in natural science determine the scope rather than the existence of
the mathematical constitution of the physical. In Sect. 8, I propose a new account of
mathematical explanation of physical facts. This account is sufficiently general so
as to be integrated into the frameworks of various existing accounts of explanation.
Accordingly, such accounts of explanation can be revised along the lines of the
proposed account to embody the idea that mathematical explanations of physical
facts highlight the mathematical constitution of physical facts and thus make the
explained facts intelligible or deepen and expand the scope of our understanding
of them. In Sect. 9, I consider four existing accounts of mathematical explanation
of physical facts. I argue that the exact nature of explanation in these accounts is
unclear and, moreover, that they are subject to the objection that mathematics only
plays a representational role, representing the physical facts that actually have the
explanatory power. I then show that these accounts can be revised along the lines of
the proposed account so as to circumvent the above objections. I conclude in Sect.
10 by briefly commenting on the implications of the mathematical constitution of
the physical for the question of the unreasonable effectiveness of mathematics in
physics.

6.3 On the Relationship Between Mathematics and Physics

There is a broad consensus among physicists and philosophers that mathematics

constitutes the language of modern physics. Poincaré expressed this view succinctly.
[O]rdinary language is too poor and too vague to express relationships so delicate, so
rich, and so precise. This, then, is a primary reason why the physicist cannot do without
mathematics; it furnishes him with the only language he can speak (Poincaré 1913, p. 141).

Lévy-Leblond (1992, p. 147) notes that the picture of mathematics as the language
of physics is ambiguous and admits two different interpretations: mathematics as
the intrinsic language of Nature, which has to be mastered in order to understand
her workings; and mathematics as a special language into which facts of Nature
need to be translated in order to be comprehended. He refers to Galileo and
Einstein as advocates of the first view, and Heisenberg as an exponent of the second
view. He does not think of these views as contradictory but rather as extreme
positions on a continuous spectrum of opinions. Lévy-Leblond explores the question
of the peculiar relation between mathematics and physics, and he contends that
the view of mathematics as a special language of nature does not provide an
answer to this question. Instead of explaining why mathematics applies so well
to physics, this view is concerned with the supposed adequacy of mathematics
as a language for acquiring knowledge of nature in general. In his view, in other
natural sciences, such as chemistry, biology, and the earth sciences, we could
talk about the relation between mathematics and the particular science as one
of application, with the implication that “we are concerned with an instrumental
relation, in which mathematics appears as a purely technical tool, external to its
6 On the Mathematical Constitution and Explanation of Physical Facts 131

domain of implementation and independent of the concepts proper to that domain.”

By contrast, he argues that mathematics plays a much deeper role in physics and that
the relation between mathematics and physics is not instrumental (ibid., p. 148). He
maintains that mathematics is internalized by physics and, like Bachelard (1965), he
characterizes the relation of mathematics to physics as that of “constitutivity” (ibid.,
p. 149).
Lévy-Leblond does not specify what he means by this relation. He notes, though,
that by mathematics being constitutive of physics, he does not mean that a physical
concept is to be identified with or reduced to its underlying mathematical concept(s).
He also does not think of this relation as implying that a physical concept is a
mathematical concept plus a physical interpretation: “the mathematical concept
is not a skeleton fleshed out by physics, or an empty abstract form to be filled
with concrete content by physics.” And, again, following Bachelard, he does not
conceive of the “distinction between a physical concept and its mathematical
characterization” as static:
mathematics . . . is a dynamical ‘way of thinking’, leading from the experimental facts to
the constitution of the physical concept (ibid., p. 149).

Further, Lévy-Leblond rejects the Platonic interpretation of the connection

between mathematics and physics, “according to which the role of the physicist
is to decipher Nature so as to reveal ‘the hidden harmony of things’ (Poincaré),
as expressed by mathematical formulas, underlying the untidy complexity of
physical facts.” For him, the mathematical constitution of the physical does not
imply “that every physical concept is the contingent materialization of an absolute
mathematical being which is supposed to express its deepest truth, its ultimate
essence” (ibid., p. 150). He takes the “mathematical polymorphism” of physics –
namely, the possibility that physical laws and concepts may be endowed with
several mathematizations – as evidence against the Platonic viewpoint and for the
dynamical nature of the mathematical constitution of the physical (ibid., pp. 150–
152).
Lévy-Leblond also rejects the view that there is a preestablished harmony
between physical and mathematical concepts. First, he draws attention to “the
physical multivalence of mathematical structures (which is the converse of the
mathematical polymorphism of physical laws), i.e., to the possibility that the same
mathematical theory may be adequate to several completely different physical
domains” (ibid., p. 152). And, following Feynman et al. (1963, vol. 2, Section 12-7),
he maintains that equations from different phenomena are so similar not because
of the physical identity of the represented systems, properties, or laws. Rather, it
is due to “a common process of approximation/abstraction” that these phenomena
are “encompassed by analogous mathematizations” (ibid., pp. 152–153). Second,
he argues that, “contrary to some rash assertions”, it is not true that the abstract
constructions elaborated by mathematics are all useful for the physicist. He contends
that, as mathematics develops on its own, the fraction of mathematical theories
and ideas that are utilized by physics is decreasing rather rapidly: while the use
132 J. Berkovitz

of mathematics in physics keeps on growing, it does so much less rapidly than the
development of mathematics that is not utilized by physics (ibid., p. 154).
Lévy-Leblond’s reflections above are intended to probe the nature of the
mathematical constitution of the physical and the special relationship between
physics and mathematics. His main conclusions are that there is no pre-established
harmony between physics and mathematics and that the singularity of physics in
its relationship to mathematics is very difficult to account for on the view that
mathematics is merely a language for representing the physical reality. Lévy-
Leblond holds that if mathematics is conceived as merely a language, it “would
necessarily have to be a universal one, applicable in principle to each and every
natural science”, and the specificity of physics “must then be treated as only a matter
of degree, and criteria must be found to distinguish it from other sciences” (ibid., p.
156). He considers two possible ways to try to explain the specificity of physics as a
matter of degree. One is the common view that physics is more advanced. The idea
here is that “more precise experimental methods, a firmer control of experimental
conditions, permit the quantitative measurement of physical notions, which are
thus transformed into numerical magnitudes” (ibid., p. 157). The other explanation
localizes
the specificity of physics in its objects of inquiry . . . It is commonly stated that physics is
“more fundamental” than the other natural sciences. Dealing with the deepest structures of
the world, it aims to bring to light the most general laws of Nature, which are implicitly
considered as the “simplest” and thus, in an Aristotelian way, as the “most mathematical”
(ibid.).

Lévy-Leblond rejects both explanations. He contends that the first explanation per
se fails to explain the privileged status of physics, and, moreover, it relies on a
very naive view of mathematics. And he claims that the second explanation implies
that “mathematicity assumes a normative character and becomes a criterion of
scientificity”, but the development of other sciences, such as chemistry, geology,
and molecular biology seems to contradict this norm. He argues that these sciences
are characterized by a clear-cut separation between their conceptual equipment and
the mathematical techniques used – a separation that allows them to maintain a
degree of autonomy even in the face of new developments in relevant branches of
physics (ibid., pp. 148, 157–158).
In the end, Lévy-Leblond fails to find an explanation for the singular relationship
between mathematics and physics. He thus opts for a radical solution according to
which physics is defined as any area of the natural sciences where mathematics is
constitutive (ibid., p. 158).
In his commentary on Lévy-Leblond’s reflections, Pitowsky (1992, p. 166)
finds Lévy-Leblond’s characterization of physics very attractive. He points out,
though, that it is incomplete, as the question arises as to what makes a given
use of mathematics constitutive rather than a mere (non-constitutive) application.
Indeed, as one could see from the above review, the questions of the nature of the
mathematical constitution of the physical and the distinction between constitutive
and non-constitutive uses of mathematics remain open. While Pitowsky shares
6 On the Mathematical Constitution and Explanation of Physical Facts 133

Lévy-Leblond’s intuition that the use of mathematics in physics is different and

deeper, he finds it difficult to provide an analytical framework for this observation.
Pitowsky also addresses the question whether Lévy-Leblond’s view of the
relationship between mathematics and physics fits in with the reductionist ethos of
science. He remarks that while “[o]ne of the central ideals in the scientific enterprise
is to try and provide a unified explanation of all natural phenomena”, the sharp lines
of demarcation that Lévy-Leblond draws “between the sciences appears to obstruct
this idea, which has had some remarkable manifestations” (ibid., pp. 166–167).
Further, Pitowsky considers Lévy-Leblond’s observation that the ratio of mathe-
matics to mathematical physics increases with time. While he finds this observation
correct, he does not think that it settles the question of whether there is a harmony
between mathematical structures and physical phenomena for two reasons (ibid., p.
163):
(a) It is impossible to foretell which portion of mathematics is relevant to physics
and which is not. There is no clear a priori distinction between pure mathematics
and mathematical physics.
(b) There is far more abstract mathematics involved in physics than one usually
assumes.
To motivate these claims, Pitowsky discusses Bolzano’s unintuitive example of
a continuous function which is nowhere differentiable (ibid., pp. 163–165). The
example was introduced at the beginning of the nineteenth century, but was ignored
for a few decades only to be resurrected and expanded by Weierstrass around the
middle of the nineteenth century.3 About 50 years later, Wiener (1923) showed, for
the probability space of the set of all possible continuous trajectories of Brownian
particles, that (with probability one) the path taken by a Brownian particle is
nowhere differentiable. As Pitowsky (ibid., p. 165) notes, “the Brownian particle
simply does not have a velocity”, and the consequent ‘pathological’ creatures of
‘continuous nowhere differentiable curves’ were instrumental in explaining why
a very large error had occurred in previous calculations. Continuous nowhere
differentiable functions and similar creatures also play a role in more recent theories
as well, such as the theory of fractals (Mandelbrot 1977). This and other examples
suggest that abstruse mathematics plays a subtle role in physical theories, which is
not sufficiently appreciated, and that it is impossible to tell a priori whether any such
piece of mathematics will figure in future theories (Pitowsky 1992, p. 165).
Pitowsky believes that such examples and the “entanglement of mathematical
structure with physical fact” suggest that “there exists no a priori principle which
distinguishes pure mathematics from applied mathematics, mathematical physics
in particular”, and that mathematical physics is much broader than it is commonly

3 Bolzano’s example is believed to have been produced in the 1830s but the manuscript was only
discovered in 1922 (Kowalewski 1923) and published in 1930 (Bolzano 1930). Neuenschwander
(1978) notes that Weierstrass presented a continuous nowhere differentiable function before the
Royal Academy of Sciences in Berlin on 18 July 1872 and that it was published first, with his
assent, in Bois-Reymond (1875) and later in Weierstrass (1895).
134 J. Berkovitz

conceived. Thus, the question of the “correspondence” between mathematics and

physics is unavoidable, and “one can understand it in a naive, realistic sense, or
defend a thesis with a Kantian flavor” (ibid., pp. 165–166).
I find Lévy-Leblond’s and Pitowsky’s reflections on the relationship between
mathematics and physics and the effectiveness of the use of mathematics in physics
intriguing. I share with them the view that, in modern physics, mathematics is
constitutive of the physical and that the use of mathematics in physics is more
successful than in other natural sciences. I disagree, though, with Lévy-Leblond’s
opinion that mathematics is only constitutive in physics. I believe that mathematics
is also constitutive in other natural sciences. Indeed, in both of the traditional
frameworks of portraying the physical as constituted by the mathematical – the
Pythagorean and the neo-Kantian – the constitution pertains to the physical broadly
conceived across the natural sciences. Yet, having a unified framework does not
entail that the natural sciences are reducible to physics.
I also share with Lévy-Leblond and Pitowsky the view that the question of
the particular success of the use of mathematics in physics is very interesting,
important, and open. Recall that Lévy-Leblond rejects two possible explanations
for the specificity of physics among the natural sciences: one is that physics is more
advanced, and the other is that the objects of physics are the simplest and most
fundamental. I don’t find his reasoning here compelling, and I believe that both
of these lines of reasoning are worth pursuing. For example, Lévy-Leblond argues
against the second explanation that it implies a false hierarchy of sciences where
physics is the most scientific and the idea that all the natural sciences are reducible
to physics. It is not clear to me, however, why the view that the objects of physics are
the simplest and most fundamental has such implications. Indeed, one may maintain
this view yet reject the suggested implications.
Finally, I share with Pitowsky the view that in modern physics there is a
“correspondence” or “harmony” between the physical and the mathematical, which
could be understood along Pythagorean or Kantian lines. I believe that such a cor-
respondence/harmony can be accounted for by the mathematical constitution of the
physical, and in the next section I will briefly review the nature of this constitution in
the context of Pythagorean and neo-Kantian frameworks. I also agree with Pitowsky
that Lévy-Leblond’s observation that the ratio of mathematics to mathematical
physics increases with time is not evidence against the correspondence/harmony
between the physical and the mathematical, and, moreover, that there is no a priori
distinction between pure and applied mathematics. And, similarly to him, I believe
that a reflection on the history of mathematical natural science could support the
absence of such a distinction.
The question of the specificity of physics among the natural sciences deserves an
in-depth consideration. Yet, for want of space, I will have to postpone its discussion
to another opportunity and focus on the constitutive role that mathematics plays in
physics and its implications for mathematical explanations of physical facts and the
question of the unreasonable effectiveness of the use of mathematics in physics. In
Sects. 5–7, I motivate the mathematical constitution of the physical and comment on
its scope. Given the mathematical constitution of the physical, in Sect. 8 I propose
6 On the Mathematical Constitution and Explanation of Physical Facts 135

a new account of mathematical explanation of physical facts, in Sect. 9 I compare it

to exiting accounts, and in Sect. 10 I briefly discuss the question of the unreasonable
effectiveness of the use of mathematics in physics. But, first, I turn to briefly review
two ways of thinking about the mathematical constitution of physical facts.

6.4 On Conceptions of Mathematical Constitution

of the Physical

There are two traditional ways of conceiving the mathematical constitution of the
physical. One is along what I shall call the ‘Pythagorean picture’. There is no
canonical depiction of this picture. In what follows, I will consider three versions of
it, due to Aristotle, Galileo, and Einstein.
Aristotle describes the Pythagorean picture in the Metaphysics, Book 1, Part 5.
[T]he so-called Pythagoreans, who were the first to take up mathematics, not only advanced
this study, but also having been brought up in it they thought its principles were the
principles of all things. Since of these principles numbers are by nature the first, and in
numbers they seemed to see many resemblances to the things that exist and come into
being – more than in fire and earth and water (such and such a modification of numbers
being justice, another being soul and reason, another being opportunity – and similarly
almost all other things being numerically expressible); since, again, they saw that the
modifications and the ratios of the musical scales were expressible in numbers; since, then,
all other things seemed in their whole nature to be modelled on numbers, and numbers
seemed to be the first things in the whole of nature, they supposed the elements of numbers
to be the elements of all things, and the whole heaven to be a musical scale and a number.
(Aristotle 1924)

In The Assayer Galileo endorses a different version of the Pythagorean picture.

In his characterization, the universe is written in the language of mathematics, the
characters of which are geometrical figures.
Philosophy is written in this grand book — I mean the universe — which stands continually
open to our gaze, but it cannot be understood unless one first learns to comprehend the
language and interpret the characters in which it is written. It is written in the language of
mathematics and its characters are triangles, circles, and other geometrical figures without
which it is humanly impossible to understand a single word of it; without these, one is
wandering a dark labyrinth. (Galileo 1623/1960)

In his lecture “On the method of theoretical physics”, Einstein also seems to
endorse a version of the Pythagorean picture. He depicts nature as the realization of
mathematical ideas.
Our experience hitherto justifies us in believing that nature is the realization of the simplest
conceivable mathematical ideas. I am convinced that we can discover by means of purely
mathematical constructions the concepts and the laws connecting them with each other,
which furnish the key to the understanding of natural phenomena. (Einstein 1933/1954)

The Pythagorean picture has also been embraced by a number of other notable
natural philosophers and physicists. While the exact nature of this picture varies
136 J. Berkovitz

from one version to another, all share the idea that mathematics constitutes all the
elements of physical reality. Physical objects, properties, relations, facts, principles,
and laws are mathematical by their very nature, so that ontologically one cannot
separate the physical from the mathematical. This does not mean that physical
things are numbers, as it may be tempting to interpret Aristotle’s characterization
in the Metaphysics and as it is frequently suggested. Rather, the idea is that
mathematics defines the fundamental nature of physical quantities, relations, and
structures. Without their mathematical properties and relations, physical things
would have been essentially different from what they are. The Pythagoreans see
the mathematical features of the physical as intrinsic to it, so that mathematics is not
regarded merely as a ‘language’ or conceptual framework within which the physical
is represented and studied.
The main challenge for the Pythagoreans is to identify the mathematical frame-
works that define the nature of the physical, especially given that both mathematics
and natural science continue to develop. The Pythagoreans may argue, though, that
it is the role of mathematics and natural science to discover these frameworks.
The second traditional way of conceiving the mathematical constitution of the
physical is along neo-Kantian lines, especially those of Herman Cohen and Ernest
Cassirer of the Marburg school. Cassirer (1912/2005, p. 97) comments that Cohen
held that
the most general, fundamental meaning of the concept of object itself, which even
physiology presupposes, cannot be determined rigorously and securely except in the
language of mathematical physics.

For example, in mechanics,

[w]hat motion “is” cannot be expressed except in concepts of quantity; understanding
these presupposes a fundamental system of a pure doctrine of Quantity. Consequently, the
principles and axioms of mathematics become the specific foundation that must be taken as
fixed in order to give content and sense to any statement of natural science about actuality.

Cassirer held that mathematics is crucial for furnishing the fundamental scaf-
folding of physics, the intellectual work of understanding, and the construction of
physical reality.
“Pure” experience, in the sense of a mere inductive collection of isolated observations, can
never furnish the fundamental scaffolding of physics; for it is denied the power of giving
mathematical form. The intellectual work of understanding, which connects the bare fact
systematically with the totality of phenomena, only begins when the fact is represented and
replaced by a mathematical symbol. (Cassirer 1910/1923, p. 147)

The idea is that “the chaos of impressions becomes a system of numbers” and
the “objective value” is obtained in “the transformation of impression into the
mathematical ‘symbol’.” The physical analysis of an object “into the totality of its
numerically constants” is a judgment in which
the concrete impression first changes into the physically determinate object. The sensuous
quality of a thing becomes a physical object, when it is transformed into a serial
determination. The “thing” now changes from a sum of properties into a mathematical
system of values, which are established with reference to some scale of comparison. Each
6 On the Mathematical Constitution and Explanation of Physical Facts 137

of the different physical concepts defines such a scale, and thereby renders possible an
increasingly intimate connection and arrangement of the elements of the given. (Ibid., p.
149)

More generally, in Cassirer’s view, the concepts of natural science are “products
of constructive mathematical procedure”. It is only when the values of physical
constants are inserted in the formulae of general laws, “that the manifold of
experiences gains that fixed and definite structure, that makes it ‘nature’.” The
reality that natural science studies is a construction, and the construction is in
mathematical terms.
The scientific construction of reality is only completed when there are found, along with the
general causal equations, definite, empirically established, quantitative values for particular
groups of processes: as, for example, when the general principle of the conservation of
energy is supplemented by giving the fixed equivalence-numbers, in accordance with which
the exchange of energy takes place between two different fields. (Ibid., p. 230)

In the neo-Kantian school, the mathematical constitution of natural science is

the outcome of a historical process. The reality that mathematical natural science
studies is a continuous serial process, and along this process the exact nature of the
mathematics that constitutes the physical evolves.
A notable example of the mathematical constitution of the physical is the role
that the calculus plays in shaping the concepts of modern physics. For example,
the concept of the mathematical limit constitutes the concept of instantaneous
velocity and more generally the concept of instantaneous change. This constitution
marks a very significant change from the concepts of instantaneous velocity and
instantaneous change before the calculus revolution. Consider, for instance, Zeno’s
arrow paradox.4 The paradox starts from the premise that time is composed of
moments, which are indivisible, and that during any such moment the arrow has
no time to move. Thus, since the arrow does not have time to move in any particular
moment, it is concluded that the arrow can never be in motion. Zeno’s paradox
suggests that in the pre-calculus era instantaneous velocity does not make sense. By
contrast, when, based on the calculus, instantaneous velocity is defined as the limit
of a series of average velocities as the time interval approaches zero, Zeno’s paradox
can be circumvented.5

4 For the arrow paradox and its analysis, see for example Huggett (1999, 2019) and references
therein.
5 Two comments: 1. The introduction of the calculus is often presented as a solution to Zeno’s

arrow paradox. I believe that this presentation is somewhat misleading. The calculus does not
really solve Zeno’s paradox. It just evades it by fiat, i. e. by redefining the concept of instantaneous
velocity. 2. For a recent example of the role that the calculus plays in constituting physical facts,
see Stemeroff’s (2018, Chap. 3) analysis of Norton’s Dome – the thought experiment that is
purported to show that non-deterministic behaviour could exist within the scope of Newtonian
mechanics. Stemeroff argues that: (i) this thought experiment overlooks the constraints that the
calculus imposes on Newton’s theory; and that (ii) when these constraints are taken into account,
Norton’s Dome is ruled out as impossible in Newtonian universes.
138 J. Berkovitz

6.5 On the Common View of How Mathematical Models

Represent Physical Reality

In modern physics, the physical is characterized in mathematical terms. Yet, the

common view is that, while mathematics provides a rich and indispensable language
for representing the physical, the physical is fundamentally non-mathematical. If
the physical is fundamentally non-mathematical, how could a mathematical model
represent adequately physical systems?
The replies to this question are often based on a widespread view that a
mathematical model represents in terms of the similarity or identity between the
mathematical structures in the model and the structures of the physical things
that the model represents. More precisely, the idea is that a mathematical model
represents in terms of an appropriate structure-preserving mapping from the math-
ematical structures that the model posits to the structures of the physical things it
represents. Chris Pincock (2004, 2011a, b, 2012, Chap. 2) calls this conception of
representation the ‘mapping account’.
For example, a model based on Newton’s three laws of motion (together with the
relevant ‘boundary conditions’) may represent the solar system. Newton’s laws of
motion are expressed in terms of mathematical relations.
First law: The net force F on an object is zero just in case its velocity v is constant:

F = 0 ⇐⇒ dv/dt = 0; (6.1)

where t denotes time and dv/dt denotes the derivative of velocity with respect to
time.
Second law: The net force F on an object is equal to the rate of change of its linear
momentum p, p = mv, where m denotes the object’s mass:

dp d(mv) dv
F = = =m . (6.2)
dt dt dt
Third law: For every action there is an equal and opposite re-action. That is, for
every force exerted by object A, FA , on object B, B exerts a reaction force, FB ,
that is equal in size, but opposite in direction:

FA = −FB . (6.3)

These mathematical relations are supposed to correspond to the physical relations

that obtain between the forces acting on the planets and the planets’ masses, linear
momenta, velocities, and accelerations, and the correspondence is achieved by a
mathematical mapping from the model to the solar system.
In its simplest form, the mapping account posits that a mathematical
model/theory represents physical systems because the mathematical structures
6 On the Mathematical Constitution and Explanation of Physical Facts 139

it posits are isomorphic to the structures of properties of, and relations between
these systems. Intuitively, the idea is that a mathematical model and properties
and relations of the physical systems it represents share the same structures. More
precisely, there is a one-to-one mapping between the mathematical structures
in the model and properties of, and relations between the represented systems.
Mathematical models/theories could also represent by means of weaker notions
of structural identity, such as partial isomorphism (French and Ladyman 1998,
Bueno et al. 2002, Bueno and French 2011, 2018), embedding (Redhead 2001),
or homomorphism (Mundy 1986). Some other accounts of how mathematical
models/theories represent take the mapping account to be incomplete. In particular,
Bueno and Colyvan (2011, p. 347) argue that, for the mapping account to get
started, we need “something like a pre-theoretic structure of the world (or at least a
pre-modeling structure of the world).” While it is common to think of mathematical
structure as a set of objects and a set of relations on them (Resnik 1997, Shapiro
1997), Bueno and Colyvan remark that “the world does not come equipped with
a set of objects . . . and sets of relations on those. These are either constructs of
our theories of the world or identified by our theories of the world.” Thus, the
mapping account requires having what they call “an assumed structure” (ibid.). Yet,
Bueno and Colyvan accept the mapping account as a necessary part of mathematical
representation of the physical world.
While the mapping account is supposed to support the common view that
mathematical models could represent physical systems even if the physical is funda-
mentally non-mathematical, it actually undermines it. If the notion of representation
is to be adequate, the notion of identity between a mathematical structure and the
physical structure it represents has to be sufficiently precise. But it is difficult to
see how the notion of identity between mathematical and physical structures could
be sufficiently precise for the purposes of modern physics if physical structures
did not have a mathematical structure. Thus, it seems that any adequate account of
mathematical representation of the physical that is based on the mapping account
(or some similar cousin of it) will have to presuppose that the represented physical
things have mathematical structures.

6.6 On the Notion of the Physical

The above argument for the mathematical constitution of the physical is focused on
accounts of representation which are based on the mapping account of mathematical
representations. There have been various objections to the mapping account (see,
for example, Frigg and Nguyen (2018) for criticism of structuralist accounts of
representation and references therein). I believe that the main idea of the above
argument can be extended to any adequate account of mathematical representation
of the physical, as any such account would have to include some aspects of the
mapping account as a necessary component. Indeed, it is difficult to see how
140 J. Berkovitz

mathematics could have the potency it has in modern natural science if mathematical
representations did not involve mapping. In any case, there is another weighty
reason for rejecting the idea that the physical is non-mathematical. It is related to the
question of the nature of the physical, and it is applicable to all accounts of scientific
representation, independently of whether they embody the mapping account.
A key motivation for thinking about the physical as non-mathematical is an
equivocation between two notions of the physical:
(i) the physical as determined by the theoretical and experimental frameworks of
modern physics and, more generally, modern natural science; and
(ii) the physical as some kind of stuff out there, the nature of which is not defined
by these frameworks.
The second notion of the physical is too vague and rudimentary to play any
significant role in most contemporary scientific applications. In particular, theories
and models in physics actually refer to the first notion of the physical, where the
physical is constituted by the mathematical in the sense that fundamental features
of the physical are in effect mathematical. Indeed, physical objects, properties,
relations, and laws in modern physics are by their very essence characterized in
mathematical terms. Yet, although the gap between the above notions of the physical
seems unbridgeable, it is common to assume without justification that our theories
and models, which embody the first notion of the physical, are about the second
notion of the physical.
The considerations above suggest that in modern natural science:
(a) The common view – that the physical is ontologically separated from the math-
ematical – fails to explain how mathematical models/theories could represent
the physical.
(b) The mapping account is an essential component of mathematical representa-
tions of the physical.
(c) The mapping account requires that physical facts have mathematical structure.
(d) The mapping account supports the idea of the mathematical constitution of
the physical, and this constitution makes sense of the mapping account as an
essential component of mathematical representations of the physical.

6.7 On the Scope of the Mathematical Constitution

of the Physical

So far, I have not considered whether the interpretation of scientific theories/models

along realist or instrumentalist lines has relevance for the question of the mathe-
matical constitution of the physical. The reason for overlooking this issue earlier
on is that the philosophical interpretative framework of theories/models in current
natural science pertains to the scope rather than the existence of mathematical
constitution of the physical. Under both realist and instrumentalist interpretations,
6 On the Mathematical Constitution and Explanation of Physical Facts 141

the physical phenomena that current theories/models account for are mathematically
constituted. The raw data that scientists collect are typically quantitative. Further,
these data often lack the systematicity, stability, and unity required for constructing
these phenomena. The phenomena that our theories/models have to answer for are
constructed from raw data in terms of various postulates and statistical inferences
which are mathematically constituted. The scope of the mathematical constitution
of physical facts beyond the phenomena depends on the interpretative framework.
If all the current theories/models in natural science were interpreted as purely
instrumental, so that they are not supposed to represent anything beyond the
phenomena, the mathematical constitution of the physical would only concern
the phenomena; for under such interpretative framework, natural science is not
supposed to represent anything beyond physical phenomena. While there are
reasons to consider parts of modern natural science as purely instrumental, it is
doubtful that there is any adequate understanding of all of it as purely instrumental.
Accordingly, it is plausible to conceive the mathematical constitution of the physical
as pertaining also to various aspects of the physical reality beyond the phenomena.
In any case, in the account of mathematical explanation of physical facts that I
now turn to propose the extent of the mathematical constitution of the physical
has implications for the scope rather than nature of mathematical explanations of
physical facts.

6.8 A Sketch of a New Account of Mathematical Explanation

of Physical Facts

In thinking about mathematical explanations of physical facts, it is important to

distinguish between:
(I) explanations in which mathematical frameworks, structures, or facts explain
physical facts; and
(II) explanations of physical facts that merely appeal to mathematical frameworks,
structures, or facts in order to represent physical facts.
The account suggested here aims at the first type of explanations. In these expla-
nations, mathematics plays an essential role, which surpasses its representational
role. While the requirement to surpass the representational role may seem trivial,
as we shall see in the next section current accounts of mathematical explanation of
physical facts struggle to meet it.
In the proposed account, mathematical explanations of physical facts are of two
related kinds. One kind of explanation is a subspecies of physical explanations
of physical facts. In a physical explanation of physical facts, physical facts are
explained by physical facts. In the proposed account, a mathematical explanation
of physical facts is a physical explanation of physical facts that focuses on the
mathematical constitution of some of the physical facts in the explanans. That
142 J. Berkovitz

is, it is a physical explanation of physical facts that highlights the mathematical

constitution of some of the physical facts in the explanans, and by highlighting
this constitution it makes the explanandum intelligible or deepens and expands the
scope of our understanding of it. The second kind of mathematical explanation of
physical facts simply highlights the mathematical constitution of physical facts and,
similarly to the first kind of explanation, by highlighting this constitution it makes
the explanandum intelligible or deepens and expands the scope of our understanding
of it.
By physical explanation of physical facts, I shall mean explanations of: how
physical facts (could) come about; how physical facts (could) influence or make
a difference to other physical facts; how physical patterns or regularities (could)
come about; how physical facts, patterns or regularities follow from, or are related
to other physical facts, patterns, or regularities; how physical principles/laws and
physical facts entail other physical facts; how physical principles/laws entail or are
related to other physical principles/laws; etc.
The proposed account is sufficiently open-ended to be based on existing accounts
of explanation, such as the D-N, causal, and unification accounts of explanation.
However, it cannot be identified with any of these accounts since it requires
that the mathematical constitution of some physical facts in the explanans be
highlighted. Given the account’s openness, it will be easier to demonstrate how
it works by analyzing the way it circumvents the difficulties that current accounts
of mathematical explanations of physical facts encounter. In the next section, I will
consider four such accounts.
Finally, recall that, in this study, the term ‘physical facts’ is used broadly to
include natural facts. The proposed account is supposed to cover mathematical
explanations of natural facts. I believe that this account could also be extended to
social facts. For, arguably, in current social science there is also a mathematical
constitution of various social facts, economic facts being a prime example. But, for
lack of space, I leave the consideration of this issue for another opportunity.

6.9 On Mathematical Explanations of Physical Facts

Recently, the literature on the mathematical explanation of physical facts has grown
steadily and various accounts have been proposed. Here I can only consider some
of them, and the choice reflects the way the current study has developed so far.
Analyzing putative examples of mathematical explanations of physical facts, I will
argue that the structure of the explanation in these accounts is unclear and that they
are susceptible to the objection that the mathematical frameworks, structures, or
facts that they appeal to play a representational rather than explanatory role. I will
then show how these accounts could be revised along the lines of the proposed
account to circumvent these challenges.
6 On the Mathematical Constitution and Explanation of Physical Facts 143

6.9.1 On a D-N Mathematical Explanation of the Life Cycle

of ‘Periodical’ Cicadas

Baker (2005, 2009, 2017) proposes a D-N-like account of mathematical explanation

of physical facts. In presenting the account, he focuses on a natural phenomenon
that is drawn from evolutionary biology: the life-cycle of the so-called ‘periodical’
cicada. Certain species of the North-American cicada share the same kind of unusual
life-cycle, where the nymph remains in the soil for a lengthy period, and then the
adult cicada emerges after either 13 or 17 years, depending on the geographical area.
“[S]trikingly, this emergence is synchronized among all the members of a cicada
species in any given area. The adults all emerge within the same few days, they
mate, die a few weeks later and then the cycle repeats itself” (Baker 2005, p. 229).
Baker’s explanation concentrates on the prime-numbered-year cicada life-cycle and
it proceeds as follows (Baker 2005, pp. 230–233, 2009, p. 614, 2017, p. 195).
Explanation of the Periodical Cicada Life-Cycle
(1) Having a life-cycle period that minimizes intersection with other (nearby/lower)
periods is evolutionarily advantageous. [biological law]
(2) Prime periods minimize intersection (compared to non-prime periods). [number
theoretic theorem]
——————————————————————————————-
(3) Hence organisms with periodic life-cycles are likely to evolve periods that are
prime. [‘mixed’ biological/mathematical law] (from (1) and (2))
(4) Cicadas in ecosystem-type E are limited by biological constraints to periods
from 14 to 18 years. [ecological constraint]
——————————————————————————————-
(5) Hence cicadas in ecosystem-type E are likely to evolve 17-year periods. (from
(3) and (4))
Baker’s core thesis is that “the cicada case study is an example of an indispensable,
mathematical explanation of a purely natural phenomenon” (Baker 2009, p. 614).
Before turning to discuss this explanation, a few remarks are in place. First, the
explanation is actually incomplete. It is based on the implicit assumption that if
something is evolutionary advantageous, it is likely to occur. This assumption is
required for (1) and (2) to imply (3). Second, in appealing to the D-N account
of explanation, Baker (2005, p. 235) broadens “the category of laws of nature to
include mathematical theorems and principles, which share commonly cited features
such as universality and necessity.” Third, whether the explanation is in fact a good
explanation of the periodical cicada life-cycle will not matter for the analysis of
the nature of Baker’s proposed D-N-like account of mathematical explanation of
physical facts.
144 J. Berkovitz

There have been various objections to the above explanation, such as that the
choice of time units (e.g. years rather than seasons) is arbitrary, that the explanation
begs the question against nominalism, and that the mathematical facts in the
explanation play a representational rather than explanatory role (Melia 2000, 2002;
Leng 2005; Bangu 2008; Baker 2009; Daly and Langford 2009; Saatsi 2011, 2016,
2018; Koo 2015). Since my main interest here is to identify the nature of Baker’s
proposed explanation, I will focus on the last objection.
Daly and Langford (2009, p. 657) argue that
[i]t is this property of the cicadas’ life-cycle duration – this periodic intersection with the
life-cycles of certain predatory kinds of organism – that plays an explanatory role in why
their life-cycle has the duration it has. Primes have similar properties, and so successfully
index the duration. But then so also do various non-primes: measuring the life-cycle in
seasons produces analogous patterns.

That is, the idea is that the durations of the life-cycles of the cicada and of the
other relevant organisms in its environment (and the forces of evolution) explain
the cicada life-cycle duration, and the number of units these durations amount to
(relative to a given measuring system) only index the durations measured. Saatsi
(2011, 2016) expresses a similar objection, proposing that Baker’s explanation
could proceed with an alternative premise to (2) and (3):
(2/3) For periods in the range of 14–18 years the intersection-minimizing period is 17. [fact
about time]

While Daly, Langford and Saatsi have nominalist sympathies, the objection that the
mathematical fact in Baker’s explanation plays only a representational role has also
been submitted by Brown (2012, p. 11) who advocates Platonism.
Baker (2017, p. 198) replies that the above nominalist perspective lacks a scope
generality: it is inapplicable to other situations in which the ecological conditions
are different. As he points out, this is not only a hypothetical point, as there are
subspecies of the periodical cicada with 13-year life cycles. Indeed, (2/3) could be
generalized into a schema (Saatsi 2011, p. 152):
(2/3)* There is a unique intersection minimizing period Tx for periods in the range [T1 ,
. . . , T2 ] years.

Yet, Baker (ibid., p. 199) notes that while (2/3)* has a more general scope, the
explanation that this schema provides is less unified and has less depth than the
original one. Further, Baker argues that the nominalist perspective also lacks topic
generality (ibid., pp. 200–208). A mathematical explanation of physical facts has
a higher topic generality if the mathematical facts that are supposed to do the
explanation could apply to explanations of other topics. For example, Baker shows
that a revised version of the explanation in (1)–(5), which meets scope generality
and topic generality, could share the same core as an explanation of why in brakeless
fixed-gear bicycles the most popular gear ratios are 46/17, 48/17, 50/17, and 46/15.
6 On the Mathematical Constitution and Explanation of Physical Facts 145

Baker argues that: (a) for an explanation to have a high level of scope and topic
generality, mathematics is indispensable; (b) the interpretation of the mathematical
facts in a mathematical explanation of physical facts as representations of physical
facts limits the topic generality of the explanation; and (c) the optimal version of the
mathematical explanation of the cicada life-cycle “has an explanatory core that is
topic general, and is not ‘about’ any designated class of physical facts, such as facts
about time, or facts about durations” (ibid., p. 201). To establish (c), he proposes a
generalized explanation of the cicada life-cycle, which I discuss below.
I agree with Baker’s observation that the nominalist perspective restricts the
level of scope generality and topic generality of explanations. But I don’t think that
his response meets the objection that, in his account, mathematics only represents
the physical facts that actually do the explanation. In what follows, I will present
the challenge for Baker’s account, first considering the original explanation of the
periodical cicada life-cycle and then the generalized version.
Baker’s explanation in (1)–(5) is ambiguous because premise (2) is ambiguous.
It could be interpreted as:
(2i) a fact about time, i.e. a physical fact; or
(2ii) a theorem about number theory, i.e. a purely mathematical fact.

For the above D-N-like explanation to be valid, it requires (2i) rather than (2ii):
since (2ii) is a pure mathematical fact, (1) and (2ii) per se do not imply (3). Thus,
it may be argued that it is the physical fact about time in (2i) that explains why
the cicadas’ life-cycle is likely to be prime, and the mathematical fact in (2ii)
only represents this physical fact; and the impression that (1) and (2ii) entail (3)
is due to an equivocation between (2i) and (2ii). The upshot is that Baker fails
to demonstrate that the explanation of the periodical cicada life-cycle above is a
mathematical explanation of a physical fact. If we interpret premise (2) as a physical
fact about the nature of time, i.e. as (2i), the D-N-like explanation in (1)–(5) is a
physical explanation of a physical fact, and it is not clear what explanatory role the
mathematical fact about prime numbers plays. If, on the other hand, we interpret (2)
as a purely mathematical fact, i.e. as (2ii), the explanation becomes either invalid
or unclear: if the explanation is supposed to be a D-N-like explanation, where
the explanandum follows deductively from the explanans, it is invalid; and if the
explanation is not intended as a D-N-like explanation, it is not clear what kind of
explanation it is.
It may be tempting to reply that the above objection does not apply to Baker’s
generalized explanation of the cicada life-cycle, where the explanatory core is not
‘about’ any designated class of physical facts because of its topic generality. Yet, as
we shall see below, the generalized explanation is still subject to the same challenge.
146 J. Berkovitz

Generalized explanation of the periodical cicada life-cycle

(M1) The lowest common multiple (LCM) of two numbers, m, n, is maximal if
and only if m and n are coprime. [pure mathematical fact]
(UC1) The gap between successive co-occurrences of the same pair of cycle
elements of two unit cycles is equal to the LCM of their respective
lengths.
————————————————————————————————–
(UC2) Hence any pair of unit cycles with periods m and n maximizes the gap
between successive co-occurrences of the same pair of cycle elements if
and only if m and n are coprime. (from M1 and UC1)
(M2) All and only prime numbers are coprime with all smaller numbers. [pure
mathematical fact]
————————————————————————————————–
(UC3) Hence, given a unit cycle, pm , of length m and a range of unit cycles, qi ,
of lengths shorter than m, pm maximizes the gap between successive
intersections with each qi if and only if m is prime. (from UC2, M2)
(1G) For periodical organisms, having a life-cycle period that maximizes the
gap between successive co-occurrences with periodical predators is
evolutionarily advantageous. [biological law]
(2G) Periodical organisms with periodical predators whose life cycles are
restricted to multiples of a common base unit can be modeled as pairs of
unit cycles.
————————————————————————————————–
(3G) Hence organisms with periodic life-cycles that are exposed to periodic
predators with shorter life-cycles, and whose life cycles are restricted to
multiples of a common base unit, are likely to evolve periods that are
prime. [‘mixed’ biological/mathematical law] (from UC3, 1G, 2G)
(4G) North American periodical cicadas fit the application conditions stated in
premise (3G).
————————————————————————————————–
(5G) Hence periodical cicadas are likely to evolve periods that are prime.
(from 3G, 4G)
(6G) Cicadas in ecosystem-type E are limited by biological constraints to
periods from 14 to 18 years. [ecological constraint]
(7G) 17 is the only prime number between 14 and 18. [pure mathematical fact]
————————————————————————————————–
(8G) Hence cicadas in ecosystem-type E are likely to evolve 17-year periods.
(from 5G, 6G, 7G)
As it is not difficult to see, in the above explanation the explanatory core –
(M1), (M2), and (UC1) – (UC3) – is topic general. Yet, similarly to the original
explanation, one could argue that for the inference from (UC3), (1G) and (2G) to
(3G) to be valid, (UC3) has to be interpreted as a proposition about physical sys-
6 On the Mathematical Constitution and Explanation of Physical Facts 147

tems – all the physical systems that satisfy the requirement of topic generality under
consideration – and that the mathematical fact in (UC3) is only a representation
of this universal physical fact. Since (UC3) follows from (UC1) and (UC2) (and
(M1) and (M2)), for the explanation to be valid, (UC1) and (UC2) also have to be
interpreted as propositions about physical systems – again, all the physical systems
that satisfy the requirement of topic generality under consideration. Thus, like in the
original explanation, the generalized explanation of the periodical cicada life-cycle
is subject to the challenge that it is either a physical explanation of a physical fact,
invalid, or unclear.
The objection above can be divided into two related objections. (α) For the
conclusion (8G) to follow from the premises in the generalized explanation of
the cicada life-cycle, (UC1) – (UC3) have to be interpreted as propositions about
physical facts. Thus, the explanation is a physical explanation of a physical fact,
and it is not clear in what sense it is a mathematical explanation of a physical fact.
(β) Like in the original explanation of the cicada life-cycle, it may be argued that
mathematics plays a representational rather than explanatory role: it represents the
physical facts that actually explains the cicada life-cycle.
The second objection is particularly compelling against those who assume that
the physical can ontologically be separated from the mathematical. It presup-
poses that the facts about physical co-occurrences can ontologically be separated
from the corresponding mathematical facts. Similarly, the objection to the orig-
inal explanation of the cicada life-cycle presupposes that the physical fact that
prime time periods minimize intersection with other (nearby/lower) periods can
ontologically be separated from the corresponding mathematical fact. But these
presuppositions are unwarranted for those who take the mathematical to constitute
the physical. The physical fact that prime time periods minimize intersection with
other (nearby/lower) periods is constituted by number theory, in general, and
its theorems about prime numbers, in particular. Accordingly, the physical fact
that prime time periods minimize intersection with other (nearby/lower) periods
cannot be separated from the corresponding mathematical fact. Likewise, facts
about physical co-occurrences are constituted by number theory, in general, and its
theorems about prime numbers, in particular. Thus, such facts cannot be separated
from the corresponding mathematical facts.
Given the mathematical constitution of the physical, Baker’s D-N-like account of
mathematical explanation of physical facts can be revised as follows to circumvent
the objections in (α) and (β). A mathematical explanation of physical facts is a phys-
ical explanation of physical facts along the D-N model in which the mathematical
constitution of some physical facts in the explanans is highlighted. The idea here
is that by highlighting the mathematical constitution of certain physical facts in
the explanans, the explanation deepens and expands the scope of the understanding
of the physical fact in the explanandum. Focusing on the original explanation of
the periodical cicada life-cycle, the revised account has two parts. First, it derives
deductively the likelihood of the prime-numbered-year life-cycle of the periodical
cicada from general biological facts, particular facts about the periodical cicada and
its eco-system, and facts about physical time. Second, it highlights the mathematical
148 J. Berkovitz

constitution of physical time, in general, and the mathematical constitution of the

fact that prime periods of physical time minimize intersection with other time
periods, in particular. The theorem of number theory in (2ii) is a reflection of
this constitution. Time is modelled as a line of real numbers and accordingly time
intervals with a length equal to a natural number are subject to the theorems of
number theory. In highlighting the mathematical constitution of these physical facts,
the explanation deepens our understanding of the curious life-cycle of the periodical
cicada in eco-system type E and expands its potential scope to the life-cycle of other
subspecies of the periodical cicada. Turning to the generalized explanation of the
periodical cicada life-cycle, in addition to showing how the likelihood of the prime-
numbered-year life-cycle of the periodical cicada follows deductively from general
biological facts, particular facts about the periodical cicada and its eco-system, and
universal physical facts about successive co-occurrences of the same pair of cycle
elements of two unit cycles, the explanation highlights the mathematical constitution
of these latter physical facts. In highlighting these facts, the explanation deepens the
understanding of the periodical cicada life-cycle in ecosystem-type E and expands
its potential scope to explanations of other subspecies of the periodical cicada –
namely, periodical cicada in other types of ecosystems – as well as to explanations
of various other natural facts.
As it is not difficult to see, the proposed revision of Baker’s account of
mathematical explanation of physical facts is not subject to the objections in (α)
and (β). Premise (2) in the original explanation of the periodical cicada life-cycle
and premises (UC1) – (UC3) in the generalized explanation of this life-cycle are
interpreted as statements about intervals of time and physical unit cycles, respec-
tively. Thus, the objection that the explanans in these explanations do not imply
their explanandum does not apply. Further, the logical structure of the explanation
is clear, and, by construction, mathematics plays an explanatory role. That is,
the second part of the original/generalized explanation of the cicada life-cycle
highlights the mathematical constitution of the physical facts in (2)/(UC1)–(UC3)
and thus deepens and expands the scope of the understanding of this life-cycle.

6.9.2 On Structural Explanation of the Uncertainty Relations

Dorato and Felline (2011, p. 161) comment that the ongoing controversy concerning
the interpretation of the formalism of quantum mechanics may explain why
philosophers have often contrasted the poor explanatory power of quantum theory
to its unparalleled predictive capacity. Yet, they claim, “quantum theory provides
a kind of mathematical explanation of the physical phenomena it is about”, which
they refer to as structural explanation. To demonstrate their claim, they present two
case studies: one involves the quantum uncertainty relations between position and
momentum, and the other focuses on quantum nonlocality.
6 On the Mathematical Constitution and Explanation of Physical Facts 149

Following Clifton (1998, p. 7),6 Dorato and Felline (2011, p. 163) hold that
we explain some feature B of the physical world by displaying a mathematical model of
part of the world and demonstrating that there is a feature A of the model that corresponds
to B, and is not explicit in the definition of the model.

The idea here is that the explanandum B is made intelligible via its structure
similarities with its formal representative, the explanans A. How do these structural
similarities render the explanandum intelligible? Dorato and Felline propose that
“in order for such a representational relation to also be sufficient for a structural
explanation, . . . we have to accept the idea that we understand the physical
phenomenon in terms of its formal representative, by locating the latter in the
appropriate model” (ibid., p. 165).
Consider, for example, a structural explanation of why position and momentum
cannot assume simultaneously definite values. In standard quantum mechanics, sys-
tems’ position and momentum are related through Fourier transforms, and Dorato
and Felline locate the explanans as the mathematical properties of these Fourier
transforms. The formal representation of the momentum (position) of a particle
is a Fourier transform of the formal representation of its position (momentum),
and Dorato and Felline propose that these Fourier transforms are required to
make intelligible the uncertainty relations between position and momentum. These
transformations are supposed to constitute the answer to the question “why do
position and momentum not assume simultaneously sharp magnitudes?”: “because
their formal representatives in the mathematical model have a property that makes
this impossible” (ibid., p 166).
Since ‘because’ here is not supposed to have a causal interpretation, the
question arises as to its exact meaning. Dorato and Felline do not address this
question explicitly. Their structural account of explanation seems to be based
on the following key ideas: (i) A structural explanation makes the explanandum
intelligible; (ii) the assumption that the properties of a physical system exemplify
the relevant parts of the mathematical model that represents it allows one to use the
properties of the latter to make intelligible the properties of the former; (iii) there
is a structure-preserving morphism from the representing mathematical model to
the represented physical fact, and this relation ensures that the represented fact can
be made intelligible by locating its representative in the mathematical model. Yet,
these ideas leave the logical structure of the explanation unclear. In particular, the
extent to which the structural explanation is different from mere representation is
unclear. Indeed, as Dorato and Felline acknowledge, one may raise the objection
that the proposed explanation is a mere translation or redescription of the physical
explanation to be given (ibid., p. 166). They concede that, in a sense, the physical
properties carry the explanatory burden. They consider, for example, a balance with
eight identical apples, five on one pan and three on the other.

6 Clifton
(1998) appropriates, with minor modifications, a definition of explanation given by
Hughes (1993).
150 J. Berkovitz

If someone explained the dropping of the pan with five apples (or the rising of the side with
three) by simply saying “5 > 3”, he/she would not have provided a genuine explanation. The
side with five apples drops because it is heavier and because of the role that its gravitational
mass has vis-à-vis the earth, not because 5 > 3!

In reply, Dorato and Felline claim that “structural explanations are not so easily
translatable into non-mathematical terms without loss of explicative power” (ibid.,
p. 167). The problem with this reply is that it does not really address the above
objection. The question under consideration is not whether to replace mathematical
explanations with non-mathematical ones. Rather, it is a question about the nature
of the mathematical explanation on offer. In particular, it is not clear in what sense
the proposed structural explanation is a mathematical explanation of physical facts
rather than a physical explanation of physical facts in which mathematics only plays
a representational role. It is not clear, for example, that, by pointing to Fourier
transforms between the functions that represent position and momentum, we are
providing a mathematical explanation of a physical fact, and not merely a physical
explanation of a physical fact in which mathematics plays a representational role.
Another aspect that is unclear in Dorato and Felline’s structural explanation is
related to the question whether the proposed pattern of explanation requires an
interpretation of the representing mathematical model. Dorato and Felline conceive
the structural explanation as providing “a common ground for understanding
[the uncertainty relations between position and momentum], independently of the
various different ontologies underlying the different interpretations of quantum
theory” (ibid., p. 165). The question arises then: How does this conception
square with the fact that their structural explanation relies on the idea that for
a mathematical model to explain properties of a physical system the physical
system has to exemplify the relevant parts of the model? The problem here is
that the quantum-mechanical formalism represents different things under different
interpretations. Consider, for example, Bohmian mechanics (Goldstein 2017 and
references therein). The functions that represent systems’ positions and momentum
in the standard interpretation of the quantum-mechanical formalism do not repre-
sent their position and momentum in Bohmian mechanics. Under this alternative
interpretation, quantum systems always have definite position and momentum, and
systems’ position and momentum do not exemplify the Fourier transform relations.
Thus, the explanation that the quantum-mechanical formal representatives – the
Fourier transforms between position and momentum – have a property that makes
it impossible for systems to have simultaneously definite position and momentum
is not correct in this case. The functions that represent systems’ position and
momentum in the standard interpretation represent in Bohmian mechanics the
range of possible outcomes of measurements of positions and momentum and their
probabilities; and the Fourier transform relations between these functions reflect
the epistemic limitation on the knowledge of systems’ position and momentum.
The upshot is that Dorato & Felline’s structural explanation cannot circumvent the
question of the interpretation of the quantum-mechanical formalism.
Given the mathematical constitution of the physical, Dorato & Felline’s struc-
tural account of explanation can be revised along the lines proposed in Sect. 8 so as
6 On the Mathematical Constitution and Explanation of Physical Facts 151

to meet the above challenges. The main idea of the revised account is that a reference
to the mathematical representatives of physical facts is explanatory if it highlights
the mathematical constitution of these facts and thus makes them intelligible. For
example, the structural explanation of why, in standard quantum mechanics, it is
impossible for position and momentum to assume simultaneously definite values
highlights the mathematical constitution of the relations between position and
momentum and thus makes the explanandum intelligible. In Bohmian mechanics,
a reference to the mathematical representatives of position and momentum in the
quantum-mechanical formalism (i.e. to the position and momentum ‘observables’)
also highlights the mathematical constitution of physical facts, but the physical facts
are not the same as in standard quantum mechanics. In Bohmian mechanics, the
highlighting is of the mathematical constitution of the limitation on simultaneous
measurements, and accordingly knowledge, of a system’s position and momentum
at any given time.7

6.9.3 On Abstract Mathematical Explanation

of the Impossibility of a Minimal Tour Across the Bridges
of Königsberg

Pincock (2007, 2011a, b, 2012, 2015) proposes an account of mathematical

explanation which he calls abstract explanation. Like most causal accounts of
explanation, abstract explanation is based on the idea of an objective dependence
relation between the explanandum and the explanans, but the notion of dependence
in play is different from that of causal dependence (Pincock 2015, p. 877). In the
case of abstract explanation, the dependence is on an abstract entity which is more
abstract than the state of affairs being explained (ibid., p. 879). Pincock has applied
this account to various examples, one of which is the explanation of the impossibility
of a ‘minimal’ tour across the bridges of Königsberg.
The citizens of eighteenth century Königsberg wished to make a minimal tour
across the city’s bridges (see Fig. 6.1), crossing each bridge exactly once and
returning to the starting point. But they failed. Euler saw that the network of bridges
and islands yields a certain abstract structure. Represented in graph theory, this
structure has edges which correspond to bridges and vertices which correspond to
islands or banks (see Fig. 6.2). Let us call a graph ‘Eulerian’ just in case it has a
continuous path from a vertex that crosses each edge exactly once and return to that
vertex (for an example of such a graph, see Fig. 6.3). Then, Euler’s (1736/1956)

7 In Bohmian mechanics, a system’s momentum is not an intrinsic property. Rather, it is a

relational property that is generally different from the system’s momentum ‘observable’. Thus,
the momentum observable does not generally reflect the system’s momentum (Dürr et al. 2013,
Chap. 7; Lazarovici et al. 2018).
152 J. Berkovitz

Fig. 6.1 The bridges of

Königsberg

Fig. 6.2 The graph of the

bridges of Königsberg: Edges
correspond to bridges and
vertices correspond to islands
or banks

conclusions about minimal tours across bridge systems can be reformulated as the
following theorem in graph theory.8
Euler’s theorem: A graph is Eulerian if and only if it contains no odd vertices; where a
vertex is called ‘odd’ just in case it has an odd number of edges leading to it.

Pincock’s proposed abstract explanation for the impossibility of a minimal tour

across the bridges of Königsberg is that such a tour is impossible because the
corresponding graph is not Eulerian. Put another way, the impossibility of a minimal
tour across the bridges of Königsberg is abstractly dependent on the corresponding

8 Since graph theory did not exist at the time, this is obviously an anachronistic account of Euler’s
reasoning. That is not a problem for the current discussion, as the aim is not to reconstruct Euler’s
own analysis but rather to consider Pincock’s account of mathematical explanation of physical
facts. For Euler’s reasoning, see Euler (1736/1956), and for a reconstruction of it, see, for example,
Hopkins and Wilson (2004).
6 On the Mathematical Constitution and Explanation of Physical Facts 153

Fig. 6.3 An example of

Eulerian graph

graph not being Eulerian. Similarly, the impossibility of a minimal tour across
this bridge system is abstractly dependent on the corresponding graph having odd
vertices.
Pincock’s motivation for an abstract explanation of the impossibility of a minimal
walk across the Königsberg bridge system is that it seems superior because it gets to
the root cause of why such a walk is impossible by focusing on the abstract structure
of this system. More generally, abstracting away from various physical facts, such
as the bridge materials and dimensions, size of the islands and river banks, etc.,
scientists can often give better explanations of features of physical systems (Pincock
2007, p. 260).
Since abstract dependence is supposed to be objective but not causal, the question
arises as to what kind of dependence it is. Pincock considers a few candidates. First,
he (2015) discusses Woodward’s (2003, p. 221) proposal that
the common element in many forms of explanation, both causal and noncausal, is that they
must answer what-if-things-had-been-different questions. . . . When a theory or derivation
answers a what-if-things-had-been-different question but we cannot interpret this as an
answer to a question about what would happen under an intervention, we may have a
noncausal explanation of some sort.

In the case of abstract explanation, the question that is answered concerns the “the
systematic relationship between the more abstract objects and their properties [the
explanans], and the more concrete objects and their properties [the explanandum].”
Pincock comments that Woodward’s proposal could be applicable to abstract
explanations “if the sort of explanatory question here could be suitably clarified”
(ibid., p. 869).
Next, Pincock (ibid., p. 869–871) discusses Strevens’ (2008) Kairetic account
of explanation. He notes that the Kairetic account “allows important explanatory
contributions from more abstract entities like mathematical objects” (ibid., p. 869).
He considers whether this account – which, like Woodward’s account, is based on
the idea of difference making, and moreover aims to reduce all explanations to
154 J. Berkovitz

causal explanations – could be a candidate for explicating the notion of abstract

dependence. Pincock rejects this route on the ground that abstract dependence
violates Strevens’ ‘cohesion’ condition; where a causal model is cohesive “if its
realizers constitute a contiguous set in causal similarity space” (Strevens 2008,
p. 104).
Pincock (2011a) also considers another notion of difference making which is
explicated in terms of comparisons across relevant range of possibilities. Applying
this notion to the explanation of why a minimal tour across the bridges of
Königsberg is impossible, his reasoning is basically as follows. Let V be a binary
variable that denotes whether all vertices of a graph are even (where a vertex is
called ‘even’ just in case it has an even number of edges leading to it), E be a binary
variable that denotes whether the graph is Eulerian, and M be a binary variable
that denotes whether there is a minimal tour across the bridges of Königsberg. The
graph of the bridges of Königsberg has odd vertices and it fails to be Eulerian, but
many similar graphs have all even vertices and each of these graphs is Eulerian.
Thus, V makes a difference for E. Further, the bridges of Königsberg fail to have a
corresponding graph that is Eulerian and they fail to have a minimal tour, but many
similar bridge systems have a corresponding graph that is Eulerian and each of these
other bridge systems has a minimal tour. Thus, E makes a difference for M.
In the above proposal, the exact nature of abstract dependence is still unclear.
In particular, Pincock does not explicate the type of similarity he has in mind, so
it is difficult to evaluate the nature of modality involved in his proposed notion
of difference making. Relatedly, it may be argued that Pincock’s reasoning does
not establish that E (V) makes a difference for M but rather that there is a
correspondence between E (V) and M, and that such a correspondence per se does
not qualify as a difference making. It is also unclear why one needs to appeal to the
notion of difference making in order to establish the modal relation between V and
E given that one could simply provide a mathematical proof of Euler’s theorem.
Finally, Pincock (2015) considers Koslicki’s (2012) notion of ontological depen-
dence as a possible route to explicating the nature of abstract dependence. One
variety of such a dependence is the following.
Constituent Dependence: An entity, , is constituent dependent on an entity (or entities),

, just in case
is an essential constituent (or are essential constituents) of . (Koslicki,
ibid., p. 205)

An example of such ontological dependence is lightning: “for lightning to occur

is just for energy to be discharged by some electrons in a certain way, and when
lightning occurs, these electrons are constituents of the lightning” (Pincock 2015,
p. 878). Pincock remarks that “it might be tempting to try to reduce abstract
dependence to constituent dependence” and “[it] is hard to argue that this cannot
be done.” But
there is one prima facie barrier that seems difficult to overcome. . . . [a] distinguishing
feature of abstract explanation and abstract dependence is that we appeal to a more abstract
entity that has a more concrete entity as an instance. . . . By contrast, in the constituent
6 On the Mathematical Constitution and Explanation of Physical Facts 155

cases, the entities that we appeal to in the explanation are constituents of the fact to be
explained. (Ibid., p. 879)

As the brief review of Pincock’s attempts to explicate abstract dependence

demonstrates, the nature of this dependence remains unclear. Further, the abstract
account of explanation is open to the objection that mathematics just plays a
representational role. The idea here is that the graph that corresponds to the bridges
of Königsberg only represents the relevant physical structure of this bridge system
which explains the impossibility of a minimal tour (Jansson and Saatsi 2019).9
This objection has more traction in view of the difficulties in explicating abstract
dependence, and it is even more pressing for those who deny the mathematical
constitution of the physical; for it is difficult to see how the impossibility of a
minimal tour – a physical fact – could depend on a mathematical object – a graph –
if the physical is fundamentally non-mathematical.
While the prospects of explicating abstract dependence in terms of Constituent
Dependence seem dim, there are other notions of ontological dependence that
escape this fate. In particular, the notions of mathematical constitution of physical
facts we reviewed in Sect. 4 are not subject to Pincock’s concern about Constituent
Dependence. When abstract dependence involves a dependence of physical facts
on mathematical structures or facts, it could be explicated in terms of these
notions of mathematical constitution. Consider, again, the bridges of Königsberg.
The impossibility of a minimal tour across the Königsberg bridge system is due
to its physical structure. This physical structure has a mathematical constitution.
Some aspects of this constitution, which concern the topology of the Königsberg
bridge system, are highlighted by graph theory, in general, and Euler’s theorem, in
particular. Thus, abstract explanation of the impossibility of a minimal tour across
the Königsberg bridge system could be conceived as highlighting the aspects of
the mathematical constitution of the physical structure of this bridge system that
render such a tour impossible. Envisaged in this way, the explanation highlights the
ontological dependence of the physical structure of the Königsberg bridge system
on a certain mathematical structure, which is expressed in terms of graph theory.
Highlighting the relevant aspects of the mathematical constitution of this physical
structure makes the impossibility of a minimal tour across the Königsberg bridge
system intelligible. It also deepens and expands the scope of our understanding by
relating this impossibility to various other cases of possible and impossible minimal
tours.

9 Vineberg (2018) suggests that the objection above does not apply to the structuralist understanding

of mathematics. It is not clear, however, how an appeal to this conception of mathematics could
help here. One may accept the structuralist rejection of mathematical objects yet argue that
mathematical structures only represent the physical structures which actually do the explanation.
156 J. Berkovitz

6.9.4 On Explanations by Constraints that Are More Necessary

than Laws of Nature

In Because without cause: Non-causal explanations in science and mathematics,

Lange (2016) considers examples of “distinctively mathematical explanations”. For
him, distinctively mathematical explanations are subspecies of a more general type
of non-causal explanation: “explanations by constraint”.
Explanations by constraint work not by describing the world’s causal relations, but rather
by describing how the explanandum arises from certain facts (“constraints”) possessing
some variety of necessity stronger than ordinary laws of nature possess. The mathematical
facts figuring in distinctively mathematical explanations possess one such stronger variety
of necessity: mathematical necessity. (Ibid., p. 10)

Lange suggests that a distinctively mathematical explanation appeals “only to

facts (including but not always limited to mathematical facts) that are modally
stronger than ordinary laws of nature, together with contingent conditions that are
contextually understood to be constitutive of the arrangement or task at issue in the
why question” (ibid., pp. 9–10). That is, in such explanations the explanans consist
of mathematically necessary facts and possibly other facts which possess some
variety of necessity stronger than that of ordinary laws of nature (the “constraints”),
as well as contingent conditions which are presupposed by the why question under
consideration. The explanation shows that “the fact to be explained could not have
been otherwise – indeed, was inevitable to a stronger degree than could result from
the action of causal powers” (ibid., pp. 5–6). Under the presupposed contingent
conditions, the explanadum arises from the “constraints”, and it is necessary in a
stronger sense than the necessity that ordinary laws of nature mandate. Thus, the
explanation works
not by describing the world’s actual causal structure, but rather by showing how the
explanandum arises from the framework that any possible physical system (whether or not
it figures in causal relations) must inhabit, where the “possible” systems extend well beyond
those that are logically consistent with all of the actual natural laws. (Ibid., pp. 30–31)

For example, the fact that a mother fails repeatedly to distribute evenly 23
strawberries among her 3 children without cutting any strawberry is explained by
the mathematical fact that 23 cannot be divided evenly by 3 (ibid., p. 6). The
explanation shows that mother’s success is impossible in a stronger sense than
causal considerations underwrite (ibid., p. 30). The explanans in this case consist
of the mathematical fact that 23 cannot be divided evenly by 3 and the contingent
facts presupposed by the why-question under consideration: namely, that there are
23 strawberries and 3 children, and the distribution is of uncut strawberries. And
the explanation shows that, under such contingent conditions, an even distribution
of the 23 strawberries among the 3 children is impossible, where the impossibility
is stronger than natural impossibility.
Saatsi (2018) poses two challenges for Lange’s account of explanation. One
challenge is that “information about the strong degree of necessity involved
[in explanation by constraint] risks being too cheap: the exalted modal aspect
6 On the Mathematical Constitution and Explanation of Physical Facts 157

of the explanandum can be communicated without doing much explaining, and

it can be grasped without having much understanding” (ibid., p. 5). Another
challenge is to pin down the difference due to which explanations-by-constraint
work so differently from causal explanations. Saatsi proposes, as an alternative to
Lange’s approach, that causal and non-causal explanations alike “can explain by
virtue of providing what-if-things-had-been-different information that captures a
dependence relation between the explanandum and the explanans.” He believes that
this counterfactual-dependence perspective suggests that, in various distinctively
mathematical explanations, at the very least “we should not hang the analysis of
explanatoriness entirely on the hook of modal ‘constraint’.” For example, “[i]f
we squeeze out, as it were, all the modal information regarding how Mother’s
predicament would differ as a function of the number of strawberries/kids, it looks
that we are left with a very shallow explanation at best, even if we fully retain the
information concerning the exalted modal status of the explanandum” (ibid., p. 6).
Lange (2016, 2018, p. 32) resists the proposal to explicate distinctively mathe-
matical explanations in terms of counterfactual dependence. He maintains that some
explanations by constraint, and more particularly some distinctively mathematical
explanations, are associated with no pattern of counterfactual dependence. Further,
he (2018, p. 33) also argues that “[s]ometimes the explanans and explanandum
of explanations by constraint do figure in patterns of counterfactual dependence,
but these patterns fail to track explanatory relations”. Yet, while Saatsi’s view that
causal and non-causal explanations could be accounted in terms of counterfactual
dependence is controversial, his objections highlight the fact that the exact nature of
Lange’s distinctively mathematical explanations is unclear. Consider, for example,
Saatsi’s question (ibid., p. 8):
Why is it that given that mass is additive, if A has the mass of 1 kg, and B has the mass of
1 kg, then the union A+B has the mass of 2 kg?

Lange (2018, p. 35) admits that “perhaps Saatsi is correct that the answer ‘Because
1+1=2’ is ‘utterly shallow’.” But he suggests that
that impression may arise from everyone’s knowing that 1+1=2 and that this fact suffices
to necessitate the explanandum, so it is difficult to see what information someone asking
Saatsi’s question might want. Furthermore, “Because 1+1=2” may not explain A+B’s
mass at all, if A and B can chemically interact when “united”. (Ibid.)

Indeed, the mathematical fact 1 + 1 = 2 and its intuitive relation to the

explanandum are very familiar. But this familiarity conceals the fact that the exact
structure of Lange’s account of distinctively mathematical explanation is unclear. It
is not clear from Lange’s account how the “because” of explanation is supposed to
work here: how a mathematical fact per se could explain a fact about the nature of
physical objects? One may argue that the fact that 1 + 1 = 2 seems to necessitate
the explanandum is due to equivocation between: (i) the mathematical fact that
1 + 1 = 2; and (ii) the universal physical fact that, for any two possible physical
objects of 1 kg, which are ‘united’ without being changed, the total mass is 2 kg.
That is, one may argue that it is the physical fact in (ii), rather than the mathematical
fact in (i), that actually does the explanation, and that the mathematical fact in
158 J. Berkovitz

(i) just represents the physical fact in (ii). Put another way, one may argue that
the intuitive appeal of Lange’s distinctively mathematical explanation is due to the
above equivocation, taking the explanation to be a physical explanation of a physical
fact – it is an explanation of a particular physical fact in terms of a corresponding
universal physical fact – in which the mathematical fact 1 + 1 = 2 only plays a
representational role.
The above objection is particularly compelling for those who maintain that the
physical is ontologically separated from the mathematical. It can be circumvented
by acknowledging the mathematical constitution of the physical. Distinctively math-
ematical explanations of physical facts could then be understood as explanations
of physical facts in which the mathematical constitution of some physical facts is
highlighted by stating the constraints that this constitution mandates. For example,
in explaining “why the union A+B has the mass of 2 kg” by pointing out that
“1 + 1 = 2”, one appeals to the universal physical fact that for any two possible
physical objects of 1 kg, A and B, which are ‘united’ without being changed, the
total mass of A + B is 2 kg. Yet, one observes that this universal physical fact is
constituted by number theory, in general, and a mathematical fact that follows from
it, 1 + 1 = 2, in particular. This constitution is highlighted by stating the constraint
1+1=2 that it mandates. In this kind of explanation, there is no need to appeal to
counterfactual dependence. Indeed, it is not clear how counterfactual dependence
could be of any help here.

6.10 Is the Effectiveness of Mathematics in Physics

Unreasonable?

It seems to be a dogma of contemporary mainstream philosophy of science that,

fundamentally, physical facts are not mathematical, and that mathematics only
provides a language for representing the physical realm, even if this language is
indispensable. Thus, the idea of mathematical explanations of physical facts may
naturally appear puzzling or even paradoxical. I argued above that the view that
the physical is ontologically separated from the mathematical fails to make sense
of the common conception of how mathematical models and theories represent
physical phenomena and reality in modern natural science. I suggested that this
conundrum could be avoided if we accept the idea that the physical is constituted
by the mathematical. I then reviewed two traditional ways of conceiving such
a constitution: the Pythagorean and the neo-Kantian. Granted the mathematical
constitution of the physical, I proposed a new account of mathematical explanation
of physical facts. In this account, there are two related kinds of explanations. One
kind of mathematical explanation of physical facts consists in: explaining physical
facts by physical facts along the lines of the D-N, causal, unification, or any other
acceptable account of explanation; and highlighting the mathematical constitution
of some of the physical facts in the explanans and thus deepening and expanding
the scope of the understanding of the explained physical facts. The second kind of
6 On the Mathematical Constitution and Explanation of Physical Facts 159

mathematical explanation of physical facts consists in highlighting the mathematical

constitution of physical facts and thus making the explained facts intelligible or
deepening and expanding the scope of our understanding of them. I also considered
four other accounts of mathematical explanation of physical facts. I argued that,
unlike the proposed account, they are open to the objections that the nature of
their explanation is unclear and that mathematics plays only a representational role,
representing the physical facts that actually do the explanation. I then suggested
that these accounts could be revised along the lines of the proposed account so as to
circumvent both challenges.
The proposed account is neutral with respect to the controversy in the philosophy
of mathematics about the ontological status of ‘mathematical objects’ (numbers,
sets, relations, functions, etc.). Further, it applies to both realist and instrumentalist
interpretations of theories/models. The interpretation of the ontological status of
theoretical terms only determines the scope of the mathematical constitution of the
physical beyond the phenomena.
In conclusion, I turn briefly to comment on how the idea of the mathematical
constitution of the physical reflects on the question whether the effectiveness of the
use of mathematics in physics is unreasonable. It is reasonable to assume that one’s
conception of mathematics and its relation to physics is important for adjudicating
this question. Let us consider then Wigner’s conception. Wigner (1960, pp. 2–3)
holds that while
the concepts of elementary mathematics and particularly elementary geometry were
formulated to describe entities which are directly suggested by the actual world, the same
does not seem to be true of the more advanced concepts, in particular the concepts which
play such an important role in physics. . . . Most more advanced mathematical concepts,
such as complex numbers, algebras, linear operators, Borel sets – and this list could be
continued almost indefinitely – were so devised that they are apt subjects on which the
mathematician can demonstrate his ingenuity and sense of formal beauty.

He (ibid.) conceives mathematics as

the science of skillful operations with concepts and rules invented just for this purpose. The
principal emphasis is on the invention of concepts. . . . The depth of thought which goes into
the formulation of the mathematical concepts is later justified by the skill with which these
concepts are used. The great mathematician fully, almost ruthlessly, exploits the domain
of permissible reasoning and skirts the impermissible. That his recklessness does not lead
him into a morass of contradictions is a miracle in itself . . . The principal point which
will have to be recalled later is that the mathematician could formulate only a handful of
interesting theorems without defining concepts beyond those contained in the axioms and
that the concepts outside those contained in the axioms are defined with a view of permitting
ingenious logical operations which appeal to our aesthetic sense both as operations and also
in their results of great generality and simplicity.

Based on this conception and consideration of the application of various “advanced

mathematical concepts” in physics, Wigner concludes that the appropriateness of
the language of mathematics for the formulation of the laws of physics is a miracle,
and he hopes that this miracle will continue in future research.
On Wigner’s conception of mathematics, the relation between the physical and
mathematical is not intrinsic. Advanced mathematics largely develops on its own
160 J. Berkovitz

and is then picked up by physicists to yield great success. It may thus be natural
to see the success of the application of the language of mathematics in physics as
surprising, mysterious, and even unreasonable. However, Wigner’s conception is
controversial (see, for example, Lützen 2011, Ferreirós 2017, Islami 2017, and ref-
erences therein). Some have argued that while Wigner’s view of mathematics seems
to have been inspired by the formalist philosophy of mathematics, this philosophy
has lost its credibility by the second half of the twentieth century (Lützen 2011,
Ferreirós 2017). Further, historically, the development of mathematics has been
entangled with that of the natural sciences, in general, and physics, in particular, and
Wigner’s conception of mathematics fails to reflect this fact. Consider, for example,
complex numbers. Wigner singles them out as one of the prime examples of most
advanced mathematical concepts that mathematicians invented for the sole purpose
of demonstrating their ingenuity and sense of formal beauty without any regard
to possible applications. Indeed, it seems that complex numbers were introduced
by Scipione del Ferro, Niccola Fontana (also known as Tartaglia, the stammerer),
Gerolamo Cardano, Ludovico Ferrari, and Rafael Bombelli in the sixteenth century
with no intent in mind to apply them. Yet, the context of the introduction was the
attempt to solve the general forms of quadratic and cubic equations, which by that
time had a long history of applications (Katz 2009). Thus, taking into account this
broader context, the claim that the introduction of complex numbers was solely for
the purpose of demonstrating ingenuity and sense of formal beauty is misleading.
That is not to argue that all the developments in mathematics were connected
to physics, neither to deny that in many cases future applications of mathematical
concepts could not be anticipated at the time of their introduction. Yet, attention to
the historical entanglement between the developments of mathematics and physics
casts doubt on the validity and scope of Wigner’s argument for the unreasonable
success of the application of mathematics in physics.
In any case, while Wigner’s view of mathematics may provide some support to
the view that the success of the application of mathematics in physics is unrea-
sonable, things are very different if we conceive the mathematical as constitutive
of the physical. In the context of such a conception, the relationship between
mathematics and physics is intrinsic. The physical is characterized in mathematical
terms. The physicist’s crude experience is formulated in precise mathematical terms,
often in statistical models of the phenomena, and accordingly the gap between the
phenomena and the theoretical models that account for them diminishes. There is
no essential gap between the concepts of elementary mathematics and geometry
and more advanced mathematical concepts. The division between applied and
theoretical mathematics is neither a priori nor fundamental. And the appropriateness
of the use of the language of mathematics in future physics is not in doubt. Thus, it is
reasonable to expect the language of mathematics to be appropriate for representing
physical reality.
Of course, such a conception of the role of mathematics in physics does not
obliterate the sense of wonder that one has with respect to the fact that “in spite
of the baffling complexity of the world, certain regularities in the events could
be discovered” (Wigner, ibid., p. 4). Yet, the reasons to conceive this wonder in
6 On the Mathematical Constitution and Explanation of Physical Facts 161

miraculous terms are not as compelling as for those who deny the mathematical
constitution of physical facts.

Acknowledgments I owe a great debt to Itamar Pitowsky. Itamar’s graduate course in the
philosophy of probability stimulated my interest in the philosophical foundations of probability
and quantum mechanics. Itamar supervised my course essay and MA thesis on the application of
de Finetti’s theory of probability to the interpretation of quantum probabilities. During my work
on this research project, I learned from Itamar a great deal about the curious nature of probabilities
in quantum mechanics. Itamar was very generous with his time and the discussions with him were
always enlightening. Our conversations continued for many years to come, and likewise they were
always helpful in clarifying and developing my thoughts and ideas. Itamar’s untimely death has
been a great loss. Whenever I have an idea I would like to test, I think of him and wish we could
talk about it. I sorely miss the meetings with him and I wish I could discuss with him the questions
and ideas considered above.
I am very grateful to the volume editors, Meir Hemmo and Orly Shenker, for inviting me
to contribute, and to Meir for drawing my attention to Itamar’s commentary on the relationship
between mathematics and physics. The main ideas of the proposed account of mathematical
explanations of physical facts were first presented in the workshop on Mathematical and Geo-
metrical Explanations at the Universitat Autònoma de Barcelona (March 2012), and I thank Laura
Felline for inviting me to participate in the workshop. Earlier versions of this paper were also
presented in the 39th, 40th, 43rd, and 46th Dubrovnik Philosophy of Science conferences, IHPST,
Paris, IHPST, University of Toronto, Philosophy, Università degli Studi Roma Tre, Philosophy,
Università degli Studi di Firenze, CSHPS, Victoria, CPNSS, LSE, Philosophy, Leibniz Universität
Hannover, ISHPS, Jerusalem, Munich Center for Mathematical Philosophy, LMU, Faculty of
Sciences, Universidade de Lisboa. I would like to thank the audiences in these venues for their
helpful comments. For discussions and comments on earlier drafts of the paper, I am very grateful
to Jim Brown, Donald Gillies, Laura Felline, Craig Fraser, Aaron Kenna, Flavia Padovani, Noah
Stemeroff, and an anonymous referee. The research for this paper was supported by SSHRC Insight
and SIG grants as well as Victoria College travel grants.

References

Aristotle. (1924). Metaphysics. A revised text with introduction and commentary by W. D. Ross (2
vols). Oxford: Clarendon Press.
Bachelard, G. (1965). L’ activité rationaliste de la physique contemporaine. Paris: Presses
Universitaires de France.
Baker, A. (2005). Are there genuine mathematical explanations of physical phenomena? Mind,
114(454), 223–238.
Baker, A. (2009). Mathematical explanations in science. British Journal for the Philosophy of
Science, 60(3), 611–633.
Baker, A. (2017). Mathematics and explanatory generality. Philosophia Mathematica, 25(2), 194–
209.
Bangu, S. (2008). Inference to the best explanation and mathematical explanation. Synthese,
160(1), 13–20.
Baron, S., Colyvan, M., & Ripley, D. (2017). How mathematics can make a difference. Philoso-
phers’ Imprint, 17(3), 1–19.
Batterman, R. (2002). Asymptotics and the role of minimal models. British Journal for the
Philosophy of Science, 53(1), 21–38.
Batterman, R. (2010). On the explanatory role of mathematics in empirical science. British Journal
for the Philosophy of Science, 61(1), 1–25.
162 J. Berkovitz

Batterman, R. (2018). Autonomy of theories: An explanatory problem. Noûs, 52(4), 858–873.

Batterman, R., & Rice, C. (2014). Minimal model explanations. Philosophy of Science, 81(3),
349–376.
Bokulich, A. (2008a). Can classical structures explain quantum phenomena? British Journal for
the Philosophy of Science, 59(2), 217–235.
Bokulich, A. (2008b). Reexamining the quantum-classical relation: Beyond reductionism and
pluralism. Cambridge: Cambridge University Press.
Bokulich, A. (2011). How scientific models can explain. Synthese, 180(1), 33–45.
Bolzano, B. (1930). Functionenlehre. In K. Rychlik (Ed.) Spisy Bernarda Bolzana Vol. 1. Prague:
Royal Bohemian Academy of Sciences.
Brown, J. R. (2012). Plantonism, naturalism, and mathematical knowledge. London: Routledge.
Bueno, O., & Colyvan, M. (2011). An inferential conception of the application of mathematics.
Nous, 45(2), 345–374.
Bueno, O., & French, S. (2011). How theories represent. British Journal for the Philosophy of
Science, 62(4), 857–894.
Bueno, O., & French, S. (2018). Applying mathematics: Immersion, inference, interpretation.
Oxford: Oxford University Press.
Bueno, O., French, S., & Ladyman, J. (2002). On representing the relationship between the
mathematical and the empirical. Philosophy of Science, 69(3), 497–518.
Cassirer, E. (1910/1923). Substanzbegriff und funktionsbegriff. Untersuchungen über die Grund-
fragen der Erkenntniskritik. Berlin: Bruno Cassirer. Translated as Substance and function.
Chicago: Open Court.
Cassirer, E. (1912/2005). Herman Cohen and the renewal of Kantian philosophy (trans. By Lydia
Patton). Angelaki, 10(1), 95–104.
Clifton, R. (1998). Scientific explanation in quantum theory. PhilSci Archive. http://philsci-
archive.pitt.edu/91/
Colyvan, M. (2001). The indispensability of mathematics. Oxford: Oxford University Press.
Colyvan, M. (2002). Mathematics and aesthetic considerations in science. Mind, 111(441), 69–74.
Daly, C., & Langford, S. (2009). Mathematical explanation and indispensability arguments.
Philosophical Quarterly, 59(237), 641–658.
Dorato, M., & Felline, L. (2011). Scientific explanation and scientific structuralism. In A. Bokulich
& P. Bokulich (Eds.), Scientific structuralism (Boston studies in the philosophy of science) (pp.
161–177). Dordrecht: Springer.
du Bois-Reymond, P. (1875). Versuch einer classification der willkürlichen functionen reeller
argumente nach ihren aenderungen in den kleinsten intervallen. Journal für die reine und
angewandte Mathematik, 79, 21–37.
Dürr, D., Goldstein, S., & Zanghì, N. (2013). Quantum physics without quantum philosophy.
Berlin: Springer.
Einstein, A. (1933/1954). On the method of theoretical physics. In A. Einstein, Ideas and opinions
(new translations and revisions by S. Bargmann), New York: Bonanza Books (1954), pp. 270–
276.
Euler, L. (1736/1956). Solutio problematis ad geometriam situs pertinentis. Commentarii
Academiae Scientiarum Imperialis Petropolitanae, 8, 128–140.
Felline, L. (2018). Mechanisms meet structural explanations. Synthese, 195(1), 99–114.
Ferreirós, J. (2017). Wigner’s “unreasonable effectiveness” in context. Mathematical Intelligencer,
39(2), 64–71.
Feynman, R., et al. (1963). The Feynman lectures in physics (Vol. 2). Reading: Addison-Wesley.
French, S., & Ladyman, J. (1998). A semantic perspective on idealization in quantum mechanics.
In N. Shanks (Ed.), Idealization IX: Idealization in contemporary physics (Pozna’n Studies in
the Philosophy of the Sciences and the Humanities) (pp. 51–73). Amsterdam: Rodopi.
Frigg, R., & Nguyen, J. (2018). Scientific representation. The Stanford Encyclopedia of philosophy
(Winter 2018 Edition). In E. N. Zalta (Ed.). URL=https://plato.stanford.edu/archives/win2018/
entries/scientific-representations/
6 On the Mathematical Constitution and Explanation of Physical Facts 163

Galileo, G. (1623/1960). The Assayer. In the controversy of the comets of 1618: Galileo Galilei,
Horatio Grassi, Mario Guiducci, Johann Kepler (S. Drake, & C. D. O’Malley, Trans.)
Philadelphia: The University of Pennsylvania Press.
Goldstein, S. (2017). Bohmian mechanics. The Stanford Encyclopedia of Philosophy (Summer
2017 Edition), E. N. Zalta (Ed.). URL=https://plato.stanford.edu/archives/sum2017/entries/
qm-Bohm/
Hopkins, B., & Wilson, R. J. (2004). The truth about Königsberg. The College Mathematics
Journal, 35(3), 198–207.
Huggett, N. (Ed.). (1999). Space from Zeno to Einstein: Classic readings with a contemporary
commentary. Cambridge, MA: MIT Press.
Huggett, N. (2019). Zeno paradoxes. The Stanford Encylopedia of Philosophy (Spring 2019
Edition), E. N. Zalta (ed.). URL= https://plato.stanford.edu/archives/win2019/entries/paradox-
zeno/
Hughes, R. I. G. (1993). Theoretical explanation. Midwest Studies in Philosophy, XVIII, 132–153.
Islami, A. (2017). A match not made in heaven: On the applicability of mathematics in physics.
Synthese, 194(12), 4839–4861.
Jansson, L., & Saatsi, J. (2019). Explanatory abstraction. British Journal for the Philosophy of
Science, 70(3), 817–844.
Katz, V. J. (2009). A history of mathematics: An introduction (3rd ed.). Reading: Addision-Wesley.
Koo, A. (2015). Mathematical explanation in science. PhD thesis, IHPST, University of Toronto.
Koslicki, K. (2012). Varieties of ontological dependence. In F. Correia & B. Schnieder (Eds.),
Metaphysical grounding: Understanding the structure of reality (pp. 186–213). Cambridge:
Cambridge University Press.
Kowalewski, G. (1923). Über Bolzanos nichtdiffrenzierbare stetige funktion. Acta Mathematica,
44, 315–319.
Lange, M. (2016). Because without cause: Non-causal explanations in science and mathematics.
Oxford: Oxford University Press.
Lange, M. (2018). Reply to my critics: On explanations by constraint. Metascience, 27(1), 27–36.
Lazarovici, D., Oldofredi, A., & Esfeld, M. (2018). Observables and unobservables in quantum
mechanics: How the no-hidden-variables theorems support the Bohmian particle ontology.
Entropy, 20(5), 116–132.
Leng, M. (2005). Mathematical explanation. In C. Cellucci & D. Gillies (Eds.), Mathematical
reasoning and heuristics (pp. 167–189). London: King’s College Publishing.
Lévy-Leblond, J.-M. (1992). Why does physics need mathematics? In E. Ullmann-Margalit (Ed.),
The scientific enterprise (Boston Studies in the Philosophy of Science) (Vol. 146, pp. 145–161).
Dordrecht: Springer.
Lützen, J. (2011). The physical origin of physically useful mathematics. Interdisciplinary Science
Reviews, 36(3), 229–243.
Lyon, A. (2012). Mathematical explanations of empirical facts, and mathematical realism.
Australasian Journal of Philosophy, 90(3), 559–578.
Mancosu, P. (2018). Explanation in Mathematics. The Stanford Encyclopedia of Philosophy
(Summer 2018 Edition), E. N. Zalta (Ed.). URL=https://plato.stanford.edu/archives/sum2018/
entries/mathematics-explanation/
Mandelbrot, B. (1977). Fractals, form, chance and dimension. San Francisco: W.H. Freeman.
Melia, J. (2000). Weaseling away the indispensability argument. Mind, 109(435), 455–479.
Melia, J. (2002). Response to Colyvan. Mind, 111(441), 75–79.
Mundy, B. (1986). On the general theory of meaningful representation. Synthese, 67(3), 391–437.
Neuenschwander, E. (1978). Riemann’s example of a continuous “nondifferentiable” function.
Mathematical Intelligencer, 1, 40–44.
Pincock, C. (2004). A new perspective on the problem of applying mathematics. Philosophia
Mathematica, 12(2), 135–161.
Pincock, C. (2007). A role for mathematics in the physical sciences. Noûs, 41(2), 253–275.
Pincock, C. (2011a). Abstract explanation and difference making. A colloquium presentation in
the Munich Center for Mathematical Philosophy, LMU, December 12, 2011.
164 J. Berkovitz

Pincock, C. (2011b). Discussion note: Batterman’s “On the explanatory role of mathematics in
empirical science”. British Journal for the Philosophy of Science, 62(1), 211–217.
Pincock, C. (2012). Mathematics and scientific representation. Oxford: Oxford University Press.
Pincock, C. (2015). Abstract explanations in science. British Journal for the Philosophy of Science,
66(4), 857–882.
Pitowsky, I. (1992). Why does physics need mathematics? A comment. In E. Ullmann-Margalit
(Ed.), The scientific enterprise (Boston Studies in the Philosophy of Science) (Vol. 146, pp.
163–167). Dordrecht: Springer.
Poincaré, H. (1913). La valeur de la science. Geneva: Editions du Cheval Ailé.
Redhead, M. (2001). Quests of a realist: Review of Stathis Psillos’s Scientific Realism: How
science tracks truth. Metascience, 10(3), 341–347.
Resnik, M. D. (1997). Mathematics as a science of patterns. Oxford: Clarendon Press.
Saatsi, J. (2011). The enhanced indispensability argument: Representational versus explanatory
role of mathematics in science. British Journal for the Philosophy of Science, 62(1), 143–154.
Saatsi, J. (2016). On the “indispensable explanatory role” of mathematics. Mind, 125(500), 1045–
1070.
Saatsi, J. (2018). A pluralist account of non-causal explanations in science and mathematics:
Review of M. Lange’s Because without cause: Non-causal explanation in science and
mathematics. Metascience, 27(1), 3–9.
Shapiro, S. (1997). Philosophy of mathematics: Structure and ontology. Oxford: Oxford University
Press.
Steiner, M. (1978). Mathematical explanations. Philosophical Studies, 34(2), 135–151.
Steiner, M. (1998). The applicability of mathematics as a philosophical problem. Cambridge, MA:
Harvard University Press.
Stemeroff, N. (2018). Mathematics, structuralism, and the promise of realism: A study of the
ontological and epistemological implications of mathematical representation in the physical
sciences. A PhD thesis submitted to the University of Toronto.
Strevens, M. (2008). Depth: An account of scientific explanation. Cambridge, MA: Harvard
University Press.
Vineberg, S. (2018). Mathematical explanation and indispensability. Theoria, 33(2), 233–247.
Weierstrass, K. (1895). Über continuirliche functionen eines reellen arguments, die für keinen
werth des letzteren einen bestimmten differentialquotienten besitzen (read 1872). In Mathema-
tische werke von Karl Weierstrass (Vol. 2, pp. 71–76). Berlin.
Wiener, N. (1923). Differential space. Journal of Mathematical Physics, 2, 131–174.
Wigner, E. (1960). The unreasonable effectiveness of mathematics in the natural sciences.
Communications in Pure and Applied Mathematics, 13(1), 1–14.
Woodward, J. (2003). Making things happen: A theory of causal explanation. Oxford: Oxford
University Press.
Chapter 7
Eerettian Probabilities, The
Deutsch-Wallace Theorem and the
Principal Principle

Harvey R. Brown and Gal Ben Porath

Chance, when strictly examined, is a mere negative word, and

means not any real power which has anywhere a being in
nature. David Hume (Hume 2008)

[The Deutsch-Wallace theorem] permits what philosophy would

hitherto have regarded as a formal impossibility, akin to
deriving an ought from an is, namely deriving a probability
statement from a factual statement. This could be called
deriving a tends to from a does. David Deutsch (Deutsch 1999)

[The Deutsch-Wallace theorem] is a landmark in decision

theory. Nothing comparable has been achieved in any chance
theory. . . . [It] is little short of a philosophical sensation . . . it
shows why credences should conform to [quantum chances].
Simon Saunders (Saunders 2020)

Abstract This paper is concerned with the nature of probability in physics, and in
quantum mechanics in particular. It starts with a brief discussion of the evolution
of Itamar Pitowsky’s thinking about probability in quantum theory from 1994
to 2008, and the role of Gleason’s 1957 theorem in his derivation of the Born
Rule. Pitowsky’s defence of probability therein as a logic of partial belief leads
us into a broader discussion of probability in physics, in which the existence
of objective “chances” is questioned, and the status of David Lewis influential
Principal Principle is critically examined. This is followed by a sketch of the work
by David Deutsch and David Wallace which resulted in the Deutsch-Wallace (DW)
theorem in Everettian quantum mechanics. It is noteworthy that the authors of

H. R. Brown ()
Faculty of Philosophy, University of Oxford, Oxford, UK
e-mail: harvey.brown@philosophy.ox.ac.uk
G. Ben Porath
Department of History and Philosophy of Science, University of Pittsburgh, Pittsburgh, PA, USA

© Springer Nature Switzerland AG 2020 165

M. Hemmo, O. Shenker (eds.), Quantum, Probability, Logic,
Jerusalem Studies in Philosophy and History of Science,
https://doi.org/10.1007/978-3-030-34316-3_7
166 H. R. Brown and G. Ben Porath

this important decision-theoretic derivation of the Born Rule have different views
concerning the meaning of probability. The theorem, which was the subject of a
2007 critique by Meir Hemmo and Pitowsky, is critically examined, along with
recent related work by John Earman. Here our main argument is that the DW
theorem does not provide a justification of the Principal Principle, contrary to the
claims by Wallace and Simon Saunders. A final section analyses recent claims to the
effect that the DW theorem is redundant, a conclusion that seems to be reinforced
by consideration of probabilities in “deviant” branches of the Everettian multiverse.

Keywords Probability · Quantum mechanics · Everett interpretation ·

Deutsch-Wallace theorem · Principal Principle · Gleason’s theorem

7.1 Introduction

Itamar Pitowsky was convinced that the heart of the quantum revolution concerned
the role of probability in physics, and in 2008 defended the notion that quantum
mechanics is essentially a new theory of probability in which the Hilbert space for-
malism is a logic of partial belief (Sect. 7.2 below). In the present paper we likewise
defend a subjectivist interpretation of probability in physics, but not restricted to
quantum theory (Sect. 7.3). We then turn (Sect. 7.4) to the important work by David
Deutsch and David Wallace over the past two decades that has resulted in the so-
called Deutsch-Wallace (DW) theorem, providing a decision-theoretic derivation
of the Born Rule in the context of Everettian quantum mechanics. This derivation
is arguably stronger than that proposed by Pitowsky in 2008 based on the 1957
Gleason theorem and further developed by John Earman, but it is noteworthy that
Deutsch and Wallace have differing views on the meaning of probability. In Sect. 7.5
we attempt to rebut the claim by Wallace and Simon Saunders – not Deutsch
– that the DW theorem provides a justification of Lewis’ influential Principal
Principle. Finally in Sect. 7.6 we analyse the 2007 critique of the DW theorem
by Meir Hemmo and Pitowsky as well as Wallaces reply. We also consider recent
claims to the effect that the theorem is superfluous, a conclusion that seems to
be reinforced by consideration of probabilities in “deviant” branches within the
Everettian multiverse.

7.2 Quantum Probability and Gleason’s Theorem

Itamar Pitowsky, in attempting to explain why the development of quantum

mechanics was a revolution in physics, argued in 1994 that at the deepest level,
the reason has to do with probability.
7 Everettian Probabilities, The Deutsch-Wallace Theorem and the Principal Principle 167

. . . the difference between classical and quantum phenomena is that relative frequencies of
microscopic events, which are measured on distinct samples, often systematically violate
some of Boole’s conditions of possible experience.1

Pitowsky’s insightful exploration of the connections between’s George Boole’s 1862

work on probability, polytopes in geometry, and the Bell inequality in quantum
mechanics is well known. In 1994, he regarded the experimental violation of the
Bell inequality as “the edge of a logical contradiction”. After all, weren’t Boole’s
conditions of possible experience what any rational thinker would come up with
who was concerned about the undeniable practical connection between probability
and finite frequencies?
Not quite. Pitowsky was clear that a strict logical inconsistency only comes
about if frequencies violating the Boolean conditions are taken from a single
sample. Frequencies taken from a batch of samples (as is the case with Bell-type
experiments) need not, for a number of reasons. The trouble for Pitowsky in 1994
was that none of the explanations for the Bell violations that had been advanced until
then in quantum theory (such as Fine’s prism model or non-local hidden variables)
seemed attractive to him.
This strikes us as more a problem, in so far as it is one, in physics than logic. At
any rate, it is noteworthy that in a 2006 paper, Pitowsky came to view the violation
of Bell-type inequalities with much more equanimity, and in particular less concern
about lurking logical contradictions. Now it is a “purely probabilistic effect” that
is intelligible once the meaning of probability is correctly spelt out in the correct
axiomatisation of quantum mechanics. What is this notion of probability?
In his 1994 paper, Pitowsky did not define it. He stated that Boole’s analysis did
not require adherence to any statement about what probability means; it is enough
to accept that in the case of repeatable (exchangeable or independent) events,
probability is manifested in, but not necessarily defined by, frequency. However in
2006, he was more committal:
. . . a theory of probability is a theory of inference, and as such, a guide to the formulation
of rational expectations.2

In the same paper, Pitowsky argued that quantum mechanics is essentially a new
theory of probability, in which the Hilbert space formalism is a “logic of partial
belief” in the sense of Frank Ramsey.3 Pitowsky now questioned Boole’s view of
probabilities as weighted averages of truth values, which leads to the erroneous
“metaphysical assumption” that incompatible (in the quantum sense) propositions
have simultaneous truth values. Whereas the 1994 paper expressed puzzlement
over the quantum violations of Boole’s conditions of possible experience by way

1 Pitowsky (1994, p. 108).

2 Pitowsky (2006, p. 4).
3 Ibid.
168 H. R. Brown and G. Ben Porath

of experimental violations of Bell-type inequalities, the 2006 paper proffered a

resolution of this and other “paradoxes” of quantum mechanics.4
We happen to have sympathy for Pitowsky’s view on the meaning of probability,
and will return to it below. For the moment we are interested rather in Pitowsky’s
2006 derivation of the Born Rule in quantum mechanics by way of the celebrated
1957 theorem of Gleason (Gleason 1957).
Gleason showed that the a measure function that is defined over the closed
subspaces of a separable Hilbert space H with dimension D ≥ 3 takes the form
of the trace of the product of two operators, one being the orthogonal projection
on the subspace in question, the other being a semidefinite trace class operator. In
Pitowsky’s 2006 axiomatisation of quantum mechanics, the closed subspaces of the
Hilbert space representing a system correspond to “events, or possible events, or
possible outcomes of experiments”. If the trace class operator in Gleason’s theorem
is read as the statistical (density) operator representing the state of the system, then
probabilities of events are precisely those given by the familiar Born Rule. But there
is a key condition in reaching this conclusion: such probabilities are a priori “non-
contextual”. That, of course, is the rub. As Pitowsky himself admitted, it is natural
to ask why the probability assigned to the outcome of a measurement B should be
the same whether the measurement is simultaneous with A or C, when A and C are
incompatible but each is compatible with B.
Pitowsky’s argument for probabilistic non-contextualism hinges on the commit-
ment to “a Ramsey type logic of partial belief” while representing the event structure
in quantum mechanics in terms of the lattice of closed subspaces of Hilbert space.
A certain identity involving closed subspaces is invoked, which when interpreted
in terms of “measurement events”, implies that the event {B = bj } (bj being the
outcome of a B measurement) is defined independently of how the measurement
process is set up (specifically, of which observables compatible with B are being
measured simultaneously).5 Pitowsky then invokes the rule that identical events
always have the same probability, and the non-contextualism result follows. (In
terms of projection operators, the probability of a projector is independent of which
Boolean sublattice it belongs to.)

4 We will not discuss here Pitowsky’s 2006 resolution of the familiar measurement problem, other
than to say that it relies heavily on his view that the quantum state is nothing more than a device
for the bookkeeping of probabilities, and that it implies that “we cannot consistently maintain that
the proposition ’the [Schrödinger] cat is alive’ has a truth value. Op. cit. p. 28. For a recent defence
of the view that the quantum state is ontic, and not just a bookkeeping device, see Brown (2019).
5 The identity is


k
  
l
 
{B = bj } ∩ {A = ai } = {B = bj } = {B = bj } ∩ {C = ci } (7.1)
i=1 i=1

where A, B and C measurements have possible outcomes a1 , a2 , . . . , ak ; b1 , b2 , . . . , br and

c1 , c2 , . . . , cl , respectively.
7 Everettian Probabilities, The Deutsch-Wallace Theorem and the Principal Principle 169

The argument is, to us, unconvincing. Pitowsky accepted that the non-
contextualism issue is an empirical one, but seems to have resolved it by fiat. (It
is noteworthy that in his opinion it is by not committing to the lattice of subspaces
as the event structure that advocates of the (Everett) many worlds interpretation
require a separate justification of probabilistic non-contextualism.6 ) Is making such
a commitment any more natural in this context than was, say, Kochen and Specker’s
ill-fated presumption (Kochen and Specker 1967) that truth values (probabilities
0 or 1) associated with propositions regarding the values of observables in a
deterministic hidden variable theory must be non-contextual?7 Arguably, the fact
that Pitowsky’s interpretation of quantum mechanics makes no appeal to such
theories – which must assign contextual values under pain of contradiction with the
Hilbert space structure of quantum states – may weaken the force of this question.
We shall return to this point in Sect. 7.4.3 below when discussing the extent to
which probabilistic non-contextualism avoids axiomatic status in the Deutsch-
Wallace theorem in the Everett picture.
John Earman (Earman 2018) is one among other commentators who also regard
Gleason’s theorem as the essential link in quantum mechanics between subjective
probabilities and the Born Rule. But he questions Pitowksy’s claim that in the light
of Gleason’s theorem the event structure dictates the quantum probability rule when
he reminds us that Gleason showed only that in the case of D > 2, and where H
is separable when D = ∞, the probability measure on the lattice of subspaces is
represented uniquely by a density operator on H iff it is countably additive (Earman
2018). And countable additivity does not hold in all subjective interpretations
of probability. (de Finetti, for instance, famously restricted probability to finite
additivity.)
In our view, the two main limitations of such Gleason-based derivations of the
Born Rule are the assumption of non-contextualism and the awkward fact that the
Gleason theorem fails in the case of D < 3, and thus can have no probabilistic
implications for qubits.8 One might also question whether probabilities in quantum
mechanics need be subjective. Indeed, Earman holds a dualist view, involving
objective probabilities as well (and which, as we shall see later, have a separate
underpinning in quantum mechanics).
The possibility of contextual probabilities admittedly raises the spectre of a
violation of the no-signalling principle, analogously to the way the (compulsory)
contextuality in hidden variable theories leads to nonlocality.9 It is noteworthy that
assuming no superluminal signalling and the projection postulate (which Pitowsky

6 Op. cit. Footnote 2.

7 For reasons to be skeptical ab initio about non-contextualism in hidden variable theories, see Bell
(1966).
8 Note that a Gleason-type theorem for systems with D ≥ 2 was provided by Busch (2003) but for

POMs (positive operator-valued measures) rather than PVMs (projection-valued measures). More
recent, stronger Gleason-type results are discussed in Wright and Weigert (2019).
9 See for example Brown and Svetlichny (1990).
170 H. R. Brown and G. Ben Porath

adopts) in the case of measurements on entangled systems, Svetlichny showed in

1998 that the probabilities must be non-contextual and used Gleason’s theorem to
infer the Born Rule.10
Prominent Everettians have in recent years defended a derivation of the Born
Rule based on decision theoretic principles, one which has gained considerable
attention. Critics often see probability as the Achilles’ Heel in the many worlds
picture, but for some Everettians (including the authors of the proof) the treatment
of probability is a philosophical triumph. It is curious, then, that the two authors,
David Deutsch and David Wallace, appear to disagree as to what the result means.
The disagreement is rooted in the question of what probability means – to which we
now turn.

7.3 The Riddle of Probability

7.3.1 Chances

The meaning of probability in the physical sciences has been a long-standing

subject of debate amongst philosophers (and to a lesser extent physicists), and it
remains controversial. Let us start with something that we think all discussants
of the notion of probability can agree on. That is the existence in Nature of
“chance processes or set-ups” that lead to more-or-less stable relative frequencies
of outcomes over the long term. Such an existence claim is non-trivial, and of
a straightforwardly empirical, and hence objective, nature. “Chances”, the term
widely used by philosophers for objective probabilities, are distilled in some way
from these frequencies, or at least connected somehow with them (see below).
The situation is somewhat reminiscent of Poincaré’s remark that time in physics
is the great simplifier. It is a non-trivial feature of the non-gravitational interactions
that choosing the “right” temporal parameter in the fundamental equations in the
(quantum) theory of each interaction – equations which contain derivatives with
respect to this parameter – results in the greatest simplification of such equations.11
There is a fairly clear sense in which the universal standard metric of time in non-
gravitational physics corresponds to the workings of Nature, even if there is nowhere
in Nature a system, apart from the Universe itself, that acts as a perfect clock, and
even if choosing a non-standard metric does not make physics impossible, just
more complicated. But whether time itself is to be reified is a moot point; our
own inclination is to deny it. Is objective probability, or chance, any different? Is
probability in physics any more of a candidate for reification than time is?

10 See Svetlichny (1998). For more recent derivations of the Born Rule based on no-signalling, see
Barnum (2003), and McQueen and Vaidman (2019). A useful critical review of derivations of the
Born Rule – including those of Deutsch and Wallace (see below) – is found in Vaidman (2020).
11 The fact that it is the same parameter in each case makes the phenomenon even more remarkable.
7 Everettian Probabilities, The Deutsch-Wallace Theorem and the Principal Principle 171

When philosophers write about objective probabilities, or “chances”, an example

that comes up repeatedly is the probability associated with the decay of a radioactive
atomic nucleus. Perhaps the most famous example of such an isotope is that of
Carbon-14, or 14 C, given its widespread use in anthropology and geology in the
dating of objects containing organic material. Its half-life is far better established
experimentally than the probability of heads for any given coin (another widely cited
example) and the decay process, being quantum, is often regarded as intrinsically
stochastic.12 Donald Gillies has argued that the spontaneity of radioactive decay
makes such probabilities “objective” rather than “artifactual”, the latter sort being
instantiated by repetitive experiments in physics, which necessarily involve human
intervention.13 Furthermore:
Any standard textbook of atomic physics which discusses radioactive elements will give
values for their atomic weights, atomic numbers, etc., and also for the probability of their
disintegrating in a given time period. These probabilities appear to be objective physical
constants like atomic weights, etc., and, like these last quantities, their values are determined
by a combination of theory and experiment. In determining the values of such probabilities
no bets at all are made or even considered. Yet all competent physicists interested in the
matter agree on certain standard values–just as they agree on the values of atomic weights,
etc. [our emphasis]14

In a similar vein, Tim Maudlin writes:

The half-life of tritium . . . is about 4499 days. Further experimentation could refine the
number . . . Scientific practice proceeds as if there is a real, objective, physical probability
density here, not just . . . degrees of belief. The value of the half-life has nothing to do with
the existence or otherwise of cognizers. . . . [our emphasis]15

David Wallace similarly states that

If probabilities are personal things, reflecting an agent’s own preferences and judgements,
then it is hard to see how we could be right or wrong about those probabilities, or how
they can be measured in the physicist’s laboratory. But in scientific contexts at least, both
of these seem commonplace. One can erroneously believe a loaded die to be fair (and thus
erroneously believe that the probability of it showing ‘6’ is one sixth); one can measure the
cross-section of a reaction or the half-life of an isotope. . . .
. . . as well as the personal probabilities (sometimes called subjective probabilities or
credences) there also appear to be objective probabilities (sometimes called chances), which
do not vary from agent to agent, and which are the things scientists are talking about when
they make statements about the probabilities of reactions and the like. [our emphasis]16

12 There are good reasons, however, for considering the source of the unpredictability in some if
not all classical chance processes as having a quantum origin too; see Albrecht and Phillips (2014).
But whether strict indeterminism is actually in play in the quantum realm is far from clear, as we
see below.
13 Gillies (2000), pp. 177, 178.
14 Gillies (1972), pp. 150, 151.
15 Maudlin (2007).
16 Wallace (2012), p. 137.
172 H. R. Brown and G. Ben Porath

In fact, Wallace goes so far as to say that denial of the existence of objective
probabilities is incompatible with a realist stance in the philosophy of science,
unless one entertains a “radical revision of our extant science”.17 At any rate,
Wallace is far from unique in denying that such probabilities can be defined in
terms of frequencies (finite or otherwise)18 so it seems to follow that the quantitative
element of reality he invokes is distinct from the presumably uncontentious,
qualitative one related to stable frequencies, referred to at the beginning of this
section.
Many philosophers have tried to elucidate what this element of reality is; the
range of ideas, which we will not attempt to itemize here, runs from the more-or-less
intuitive notion of “propensity” to the abstract “best system” approach to extracting
probabilities from frequencies in the Humean mosaic (or universal landscape of
events in physics, past, present and future).19 The plethora of interpretations on the
part of philosophers has a curious feature: the existence of chance is widely assumed
before clarification is achieved as to what it is! And the presumption is based on
physics, or what is taken to be the lesson of physics.20

7.3.2 Carbon-14 and the Neutron

Until relatively recently, the half-life (or 0.693 of the mean lifetime) of 14 C (which
by definition is a probabilistic notion21 ) was anomalous from a theoretical point

17 Op. cit., p. 138. It should be noted, however, that although Wallace claims that the certain

probabilities in statistical mechanics must be objective, he accepts that such a claim is hard to
make sense of without bringing in quantum mechanics (see Wallace 2014). For further discussion
of Wallace’s views on probability in physics, see (Brown, 2017).
18 For discussion of the problems associated with such a definition, see, e.g., Saunders (2005),

Greaves and Myrvold (2010), Wallace (2012), p. 247, and Myrvold, W. C., Beyond Chance and
Credence [unpublished manuscript], Sect. 3.2.
19 This last approach is due to David Lewis (1980); an interesting analysis within the Everettian

picture of a major difficulty in the best system approach, namely “undermining”, is found in
Saunders (2020). It should also be recognised that the best system approach involves a judicious
almagam of criteria we impose on our description of the world, such as simplicity and strength;
the physical laws and probabilities in them are not strict representations of the world but rather our
systematization of it. Be that as it may, Lewis stressed that whatever physical chance is, it must
satisfy his Principal Principle linking it to subjective probability; see below.
20 Relatively few philosophers follow Hume and doubt the existence of chances in the physical

world; one is Toby Handfield, whose lucid 2012 book on the subject (Handfield 2012) started out
as a defence and turned into a rejection of chances. Another skeptic is Ismael (1996); her recent
work (Ismael 2019) puts emphasis on the kind of weak objectivity mentioned at the start of this
section. See also in this connection recent work of Bacciagaluppi, whose view on the meaning of
probability is essentially the same as that defended in this paper: (Bacciagaluppi, 2020), Sect. 10.
21 The half-life is not, pace Google, the amount of time needed for a given sample to decay by

one half! (What if the number of nuclei is odd?) It is the amount of time needed for any of the
nuclei in the sample to have decayed with a probability of 0.5. Of course this latter definition will
7 Everettian Probabilities, The Deutsch-Wallace Theorem and the Principal Principle 173

of view. It is many orders of magnitude larger than both the half-lives of isotopes
of other light elements undergoing the same decay process, and what standard
calculations for nucleon-nucleon interactions in Gamow-Teller beta decay would
indicate.22
Measurements of the half-life of 14 C have been going on since 1946, involving
a number of laboratories worldwide and two distinct methods. Results up until
1961 are now discarded from the mix; they varied between 4.7 and 7.2 K years.
The supposedly more accurate technique involving mass-spectrometry has led to
the latest estimate of 5700 ± 30 years (Bé and Chechev 2012). This has again
involved compiling frequencies from different laboratories using a weighted average
(the weights depending on estimated systematic errors in the procedures within the
various laboratories).
In the case of the humble neutron, experimental determination of its half-life
is, as we write, in a curious state. Again, there are two standard techniques, the so-
called ‘beam’ and ‘bottle’ measurements. The latest and most accurate measurement
to date using the bottle approach estimates the lifetime as 877.7 ± 0.7 s (Pattie
2018), but there is a 4 standard deviations disagreement with the half-life for neutron
beta decay determined by using the beam technique which is currently not well
understood.
It should be clear that measurements involve frequencies of decays in all these
cases. The procedures might be compared to the measurement of, say, the mass of
the neutron. As Maudlin says, further experimentation could refine the numbers.
But for those who (correctly) deny that probabilities can be defined in terms of
finite frequencies, the question obviously arises: in what precise sense does taking
measurements of frequencies constitute a measurement of probability (in the present
case, half-life)? Is Gillies right that nothing like a bet is being made?
The difference between measuring mass or atomic weights, say, and half-life
is not that one case involves complicated, on-going interactions between theory and
experiment and the other doesn’t. The essential difference is that for the latter, if it is
thought to be objective, no matter how accurate the decay frequency measurements
are and how many runs are taken, the resulting frequencies could in principle all be a
fluke, and thereby significantly misleading. That is the nature of chance procedures,
whether involving isotopes or coins. Here is de Finetti, taken out of context:
It is often thought that these objections [to the claim that probability theory and other exact
sciences] may be escaped by making the relations between probabilities and frequencies
precise is analogous to the practical impossibility that is encountered in all the experimental
sciences of relating exactly the abstract notions of the theory and the empirical realities. The
analogy is, in my view illusory . . . in the calculus of probability it is the theory itself which
obliges us to admit the possibility of all frequencies. In the other sciences the uncertainty

be expected to closely approximate the former when the sample is large, given the law of large
numbers (see below).
22 In 2011, (Maris et al. 2011) showed for the first time, with the aid of a supercomputer, that

allowing for nucleon-nucleon-nucleon interactions in a dynamical no-core shell model of beta

decay explains the long lifetime of 14 C.
174 H. R. Brown and G. Ben Porath

flows indeed from the imperfect connection between the theory and the facts; in our case,
on the contrary, it does not have its origin in this link, but in the body of the theory itself
. . . .23

In practice, of course, given “enough” runs, the weighted averages of the

frequencies are taken as reliable guides to the “objective” half-life, given the
(weak) law of large numbers. It would be highly improbable were such averaged
frequencies to be significantly different from the “real” half-life (assuming we trust
our experimental methods). Note first that this meta-probability is in the nature of
a belief, or “credence” on the part of the physicist; it is not based on frequencies
on pain of an endless regress. So this is a useful reminder that in physics whether
or not we need to appeal to the existence of objective chances, we certainly need
to appeal in an important way to a subjective, or personal notion of probability,
even in quantum mechanics. And objectivists, at least, are effectively betting – that
the observed frequencies are “typical”, so as to arrive at something close to the
“objective” half-life.24

23 de Finetti (1964), p. 117. de Finetti’s argument, as Gillies (2000) stresses on pages 103 and
159, was made in the context of the falsifiability of probabilistic predictions, even in the case
of subjective probabilities. But the argument also holds in the case of inferring “objective”
probabilities by way of frequencies.
24 Saunders puts the point succinctly:

Chance is measured by statistics, and perhaps, among observable quantities, only by

statistics, but only with high chance. (Saunders 2010)
This point is repeated in Saunders (2020, Sect. 7.2). To repeat, this “high chance” translates
into subjective confidence; Saunders, like many others, believes that chances and subjective
probabilities are linked by way of the “Principal Principle” (see below). Wallace’s 2012 version
of the law of large numbers explicitly involves both “personal” and objective probabilities. His
personal prior probabilities Pr(·|C) are conditional on the proposition expressing all of his current
beliefs, denoted by C. Xp denotes the hypothesis that the objective probability of En is p, where
as usual it is assumed that p is an iid. Then, Wallace claims, the Principal Principle ensures that his
personal probability of the proposition Yi , that heads occurs in the ith repetition of the random
process, is p: symbolically Pr(Yi |Xp &C) = p. Given the iid assumption for p and the usual
updating rule for personal probabilities, it follows from combinatorics that the personal probability
Pr(KM |Xp &C) is very small unless p = M/N , where KM is the proposition that the experiment
results in M heads out of N . Wallace concludes that
. . . as more and more experiments are carried out, any agent conforming to the Principal
Principle will become more and more confident that the objective probability is close to the
observed relativity frequency. (Wallace 2012, p. 141.)
Note that there are versions of the “objective” position that avoid the typicality assumption
mentioned above, such as that defended in Gillies (2000), Chap. 7, and (Greaves and Myrvold
2010). For more discussion of the latter, see Sect. 2.3 (ii) below.
7 Everettian Probabilities, The Deutsch-Wallace Theorem and the Principal Principle 175

7.3.3 The Law of Large Numbers

(i) Whether we are objectivists, subjectivists or dualists about probability, for such
processes as the tossing of a bent coin, or the decay of Carbon-14, we will need to be
guided by experience in order to arrive at the relevant probability. For the objectivist,
this is letting relative frequencies uncover an initially unknown objective probability
or chance. For the subjectivist, it is using the standard rule of updating a (suitably
restricted) subjective prior probability in the light of new statistical evidence, subject
to a certain constraint on the priors.
Suppose we have a chance set-up, and we are interested in the probability of a
particular outcome (say heads in the case of coin tossing, or decay for the nucleus
of an isotope within a specified time) in the n + 1th repetition of the experiment,
when that outcome has occurred k times in the preceding n repetitions. For the
objectivist, the chance p at each repetition (“Bernoulli trial”) is typically assumed
to be an identically distributed and independent distribution (iid), which means that
the objective probability in question is p, whatever the value of M is. But when p
is unknown, we expect frequencies to guide us. It is their version of the (weak) law
of large numbers that allows the objectivists to learn from experience.25
The celebrated physicist Richard Feynman defined the probability of an event as
an essentially time-asymmetric notion, viz. our estimate of the most likely relative
frequency of the event in N future trials.26 This reliance on the role of the estimator
already introduces an irreducibly subjective element into the discussion.27 And note
how Feynman describes the experimental determination of probabilities in the case
of tossing a coin or a similar “chancy” process.
We have defined P (H ) = NH /N [where P (H ) is the probability of heads, and NH  is
the expected number of heads in N tosses]. How shall we know what to expect for NH ? In
some cases, the best we can do is observe the number of heads obtained in large numbers
of tosses. For want of anything better, we must set NH  = NH (observed). (How could
we expect anything else?) We must understand, however, that in such a case a different
experiment, or a different observer, might conclude that P (H ) was different. We√would
expect, however, that the various answers should agree within the deviation 1/2 N [if
P (H ) is near one-half]. An experimental physicist usually says that an “experimentally
determined” probability has an “error”, and writes

NH 1
P (H ) = ± √ (7.2)
N 2 N

There is an implication in such an expression that there is a “true” or “correct” probability

which could be computed if we knew enough, and that the observation may be in “error”

25 Thisis not to say that the law of large numbers allows for probabilities to be defined in terms of
frequencies, a claim justly criticised in Gillies (1973), pp. 112–116.
26 Feynman et al. (1965), Sect. 6.1.
27 See Brown (2011). In so far as agents are brought into the picture, who remember the past and

not the future, probability requires the existence of an entropic arrow of time in the cosmos; see in
this connection (Handfield 2012, chapter 11), Myrvold (2016) and (Brown, 2017), Sect. 7.
176 H. R. Brown and G. Ben Porath

due to a fluctuation. There is, however, no way to make such thinking logically consistent.
It is probably better to realize that the probability concept is in a sense subjective, that it
is always based on uncertain knowledge, and that its quantitative evaluation is subject to
change as we obtain more information.28

Now how far this position takes Feynman away from the objectivist stance is perhaps
debatable; note that Feynman does not refer here to quantum indeterminacies. But
other physicists have openly rejected an objective interpretation of probability; we
have in mind examples such as E. T. Jaynes (1963), Frank Tipler (2014), Don Page
(1995) and Jean Bricmont (2001). In 1995, Page memorably compared interpreting
the unconscious quantum world probabilistically to the myth of animism, i.e.
ascribing living properties to inanimate objects.29
(ii) It is indeed worth reminding ourselves of the fact that the brazen subjectivists
can make a good – though far from uncontroversial – case for the claim that their
practice is consistent with standard experimental procedure in physics.
The subjectivist specifies a prior subjective probability for such a probability
(presumably conditional on background knowledge of some kind) and appeals to
Bayes’ rule for updating probability when new evidence (frequencies) accrues.
Following the work principally of de Finetti, the crucial assumption that replaces
the iid condition for objectivists is the constraint of “exchangeability” on the
prior probabilities, which opens the door to learning from the past. de Finetti’s
1937 representation theorem based on exchangeability leads to an analogue of the
Bernoulli law of large numbers.30 In the words of Brian Skyrms,
de Finetti . . . looks at the role that chance plays in standard statistical reasoning, and argues
that role can be fulfilled perfectly well without the metaphysical assumption that chances
exist. (Skyrms 1984)

A detailed discussion of the role of the de Finetti representation theorem in

physics was provided in 2010 by Greaves and Myrvold, which appears to question
de Finetti’s own purely subjectivist interpretation of the theorem. Consider an

28 Feynman et al. (1965), Sect. 6.3. We are grateful to Jeremy Steeger for bringing this passage to
our attention. For his careful defence of a more objective variant of Feynman’s notion of probabil-
ity, see (Steeger, ’The Sufficient Coherence of Quantum States’ [unpublished manuscript]).
29 See Page (1995). Page has also defended the view that truly testable probabilities are limited to

conditional probabilities associated with events defined at the same time (and perhaps the same
place); see Page (1994). This view is motivated by the fact that strictly speaking we have no direct
access to the past or future, but only present records of the past, including of course memories.
The point is well-taken, but radical skepticism about the veracity of our records/memories would
make science impossible, and as for probabilities concerning future events conditional on the past
(or rather our present memories/records of the past) they will be testable when the time comes,
given the survival of the relevant records/memories. So it is unclear whether recognition of Page’s
concern should make much difference to standard practice.
30 de Finetti (1964); for a review of de Finetti’s work see Galavotti (1989). A helpful introduction

to the representation theorem is given by Gillies (2000, chapter 4); in this book Gillies provides a
sustained criticism of the subjectivist account of probabilities.
7 Everettian Probabilities, The Deutsch-Wallace Theorem and the Principal Principle 177

agent looking for the optimal betting strategy for sequences of a relevant class of
experimental setups. Then
. . . when he conditionalizes on the results of elements of the sequence, he learns about what
the optimal strategy is, and he is certain that any agent with non-dogmatic priors on which
the sequence of experiments is exchangeable will converge to the same optimal strategy.
If this is not the same as believing that there are objective chances, then it is something
that serves the same purpose. Rather than eliminate the notion of objective chance, we have
uncovered, in [the agent’s] belief state, implicit beliefs about chances, – or, at least, about
something that plays the same role in his epistemic life. . . .
. . . like it or not, an agent with suitable preferences acts as if she believes that there
are objective chances associated with outcomes of the experiments, about which she can
learn, provided she is non-dogmatic. . . . There may be more to be said about the nature and
ontological status of such chances, but, whatever more is said, it should not affect the basic
picture of confirmation we have sketched.31

Note that in explicitly leaving the question of the “nature and ontological status” of
chances open, Greaves and Myrvold’s analysis appeals only to the intersubjectivity
of the probabilistic conclusions drawn by (non-dogmatic) rational betting agents
who learn from experience according to Bayesian updating. Whatever one calls the
resulting probabilities, it is questionable whether they are out there in the world
independent of the existence of rational agents, and playing the role of elements of
reality seemingly demanded by Gillies, Maudlin and Wallace as spelt out in Sect. 2.1
above.32,33
(iii) In anticipation of the discussion of the refutability of probabilistic assertions
in science below (Sect. 7.6), it is worth pausing briefly to consider an important
feature of subjectivism in this sense. A subjective probability for some event, either
construed as a prior, or conditional on background information (say relevant past
frequencies), is not open to revision. Such a probability is immutable; it is not
to be “repudiated or corrected”.34 The same holds for the view that conditional
probabilities are rational inferences arising from incomplete information. What the
introduction of relevant new empirical information does is not to correct the initial
probability but replace it (by Bayesian updating) with a new one conditional on the
new (and past, if any) evidence, which in turn is immutable.35

31 Greaves and Myrvold (2010), Sect. 3.3.

32 Itis noteworthy that in Footnote 4 (op. cit.), Greaves and Myrvold entertain the view that “it is
via [the Principal Principle] that we ascribe beliefs about chances to the agent”. This Principle is
the topic of the next section of our paper, in which we report Myrvold’s more recent claim that it
is not about chances at all.
33 Greaves and Wallace also take issue (p. 22) with the claim made in the final sentence of Sect. 2.2

above that objectivists need to introduce the substantive additional assumption that the statistical
data is “typical”. As far as we can see, this is again because the notion of chance they are
considering is ontologically neutral.
34 See de Finetti (1964, pp. 146, 147).
35 Gillies (2000, pp. 74, 83, 84) and Wallace (2012, p. 138), point out that de Finetti’s scheme

of updating by Bayesian conditionalisation yields reasonable results only if the prior probability
function is appropriate. If it isn’t, no amount of future evidence can yield adequate predictions
based on Bayesian updating. This a serious concern, which will not be dealt with here (see in this
178 H. R. Brown and G. Ben Porath

(iv) We note finally that, as Myrvold has recently stressed, there is a tradition
dating back to the nineteenth century of showing another way initial credences
can “converge”, which does not involve the accumulation of new information. The
“method of arbitrary functions”, propounded in particular by Poincaré, involves
dynamical systems in which somewhat arbitrary prior probability distributions are
“swamped” by the physics, which is to say that owing to the dynamics of the system
a wide range of prior distributions converge over time to a final distribution. It is
noteworthy that Myrvold accepts that the priors can be epistemic in nature, but
regards the final probabilities as “hybrid”, neither epistemic nor objective chances.
He calls them epistemic chances.36

7.3.4 The Principal Principle

If one issue has dominated the philosophical literature on chance in recent decades –
and which will be relevant to our discussion of the Deutsch-Wallace theorem below
– it is the Principal Principle (henceforth PP), a catchy term coined and made
prominent by David Lewis in 1980 (Lewis 1980), but referring to a longstanding
notion in the literature.37
This has to do with the second side of what Ian Hacking called in 1975 the Janus-
faced nature of probability,38 or rather its construal by objectivists. The first side
concerns the issue touched on above, i.e. the estimation of probabilities on the basis
of empirical frequencies. Papineau (1996) called this the inferential linkbetween
frequencies and objective probabilities (chances). The second side concerns what
we do with these probabilities, or why we find them useful. Recall the famous
dictum of Bishop Butler that probabilities are a guide to life. Objective probabilities
determine, or should determine, our beliefs, or “credences”, about the future (and
less frequently the past) which if we are rational we will act on. Following Ramsey,
it is widely accepted in the literature that this commitment can be operationalised,
without too much collateral damage, by way of the act of betting.39

connection Greaves and Myrvold 2010). Note that it seems to be quite different from the reason
raised by Wallace (see Sect. 2.1 above) as to why the half-life of a radioactive isotope, for example,
must be an objective probability. See also Sect. 7.5.1 below.
36 Myrvold, W. C., Beyond Chance and Credence [unpublished manuscript], Chap. 5.
37 See Strevens (1999). In our view, a refreshing exception to the widespread philosophical interest

in the PP is Gillies’ excellent 2000 book (Gillies 2000), which makes no reference to it.
38 Hacking (1984), though see Gillies (2000, pp. 18, 19) for references to earlier recognition of

the dual nature of probability. Note that Wallace has argued that the PP grounds both sides; see
Footnote 23 above.
39 See Gillies (2000, chapter 4).
7 Everettian Probabilities, The Deutsch-Wallace Theorem and the Principal Principle 179

Papineau (1996) called the connection between putative objective probabilities

(chances) and credences the decision-theoretic link. More famously, it is expressed
by way of the Principal Principle; here is a simple version of it:
For any number x, a rational agent’s personal probability, or credence, of an event E
conditional on the objective probability, or chance, of E being x, and on any other accessible
background information, is also x.40

Now even modest familiarity with the literature reveals that the status of the PP
is open to dispute. According to Ned Hall (2004), it is an analytic truth (i.e. true
by virtue of the meaning of the terms “chance” and “credence”) though this view is
hardly consensual. The weaker position that chance is implicitly defined by the PP
has been criticised by Carl Hoefer.41 Amongst other philosophers who think the PP
requires justification, Simon Saunders considers that the weakness in the principle
has to do with the notoriously obscure nature of chances:
Failing an account of what objective probabilities are, it is hard to see how the [PP] could
be justified, for it seems that it ought to be facts about the physical world that dictate our
subjective future expectations. (Saunders 2005)

It is precisely the Deutsch-Wallace theorem that Saunders regards as providing the

missing account of what chances are, as we shall see. If there is a difference between
his view and that of David Wallace, it is that the latter sees the prima facie difficulty
with the PP as more having to do with the its inferential form than the clarification
of one of its components (chances). Wallace argues that justification of the rational
inference from chances to credences is generically hard to come by, unless by way
of the principles of Everettian physics and decision theory (of which more below).42
In the absence of such principles, Wallace’s position seems close to that of Michael
Strevens, who in 1999 wrote:
. . . in order to justify [the PP], it is not enough to simply define objective probability as
whatever makes [the PP] rational. In addition, it must be shown that there is something in
this world with which it is rational to coordinate subjective probabilities. . . .
As with Humean induction, so with probability coordination: we cannot conjure a
connection between the past and the future, or between the probabilistic and the non-
probabilistic, from mathematics and deductive logic alone.43

Whether the Deutsch-Wallace theorem is the ultimate conjuring trick will be

examined shortly. We finish this brief and selective survey of interpretations of
the PP with mention of the striking view of Wayne Myrvold, who, like Strevens,

40 This version is essentially the same as that in Wallace p. 140. Note that Wallace here (i) does
not strictly equate the PP with the decision-theoretic link and (ii) sees PP as underpinning not
just the decision-theoretic link but also the inferential link (see Footnote 24 above). We return to
point (i) below; we are also overlooking here the usual subtleties involved with characterising
the background information as “accessible”. For a more careful discussion of the PP, see
(Bacciagaluppi 2020).
41 Hoefer (2019), Sect. 1.3.4.
42 Wallace (2012), Sects. 4.10, 4.12.
43 Strevens (1999). See also (Hoefer, 2019, Sects. 1.2.3 and 1.3.4).
180 H. R. Brown and G. Ben Porath

wonders how chances could possibly induce normative constraints on our credences.
Myrvold cuts the Gordian Knot by claiming that the PP is concerned rather with our
beliefs about chances – even if there are no non-trivial chances in the world.
PP captures all we know about credences about chances, since it is not really about
chances.44

Can there be any doubt that the meaning and justification of the PP are issues that
are far from straightforward?
Subjectivists can of course watch the whole tangled debate with detachment,
if not amusement: if there are no chances, nor credences about chances, the
PP is empty. And Papineau’s two links collapse into one: from frequencies to
credences.45 But now an obvious and old question needs to be addressed: can
credences sometimes be determined without any reference to frequencies?

7.4 Quantum Probability Again

7.4.1 The Principle of Indifference

It has been recognised from the earliest systematic writings on probability that
symmetry considerations can also play an important role in determining credences
in chance processes. In fact, in what is often called the “classical” theory of
probability, which, following Laplace, was dominant amongst mathematicians for
nearly a century,46 the application of the principle of indifference (or insufficient
reason) based on symmetry considerations is the grounding of all credences in
physics and games of chance. The closely related, so-called “objective” Bayesian
approach, associated in particular with the name of E. T. Jaynes, likewise makes
prominent use of the principle of indifference.
David Wallace voices the opinion of many commentators when he describes this
principle as problematic.47 It is certainly stymied in the case of chance processes
lacking symmetry (such as involving a biased coin) and it is widely accepted that
serious ambiguities arise when the sample space is continuous.48 But what can be
wrong with use of the principle when assigning the probability P (H ) = 1/2 (see
Sect. 7.3.3) in the case of a seemingly unbiased coin? And are we not obtaining in

44 Myrvold, ‘Beyond Chance and Credence’ [unpublished manuscript], Sect. 2.5.

45 See Brown (2011). In this paper it is briefly argued that the related philosophical “problem of
induction” should be seen as a pseudo-problem.
46 See Gillies (2000, chapter 2) for a useful introduction to the classical theory.
47 See Wallace (2012, section 4.11).
48 A lucid discussion is found in Gillies (2000, pp. 37–48), which contains an insightful critique of

Jaynes’ defence of the principle of indifference in physics – and conceding that the principle has
been successfully applied in a number of cases in physics.
7 Everettian Probabilities, The Deutsch-Wallace Theorem and the Principal Principle 181

this case a probabilistic credence purely on the basis of a principle of rationality,

deriving a ‘tends to’ from a ‘does’, in the words of David Deutsch?
Not according to Wallace. He points to the fact that in the case of a classical
(deterministic) chance process, or even a genuinely stochastic process, the sym-
metry (if any) must ultimately be broken when any one of the possible outcomes
is observed. The obvious culprit? Either suitably randomised initial conditions, or
stochasticity itself.
. . . since only one outcome occurs, something must break the symmetry – be it actual
microconditions of the system, or the actual process that occurs in a stochastic process.
Either way, we have to build probabilistic assumptions into the symmetry-breaking process,
and in doing so we effectively abandon the goal of explicating probability.49

But if, in a branching Everettian universe where, loosely speaking, everything that
can happen does happen, a derivation of credences can be given on the basis of
rationality principles that feature the symmetry properties of the wavefunction, no
such symmetry need be broken, and a reductive analysis of probability is possible.
It is precisely this feat that the Deutsch-Wallace theorem achieves.

7.4.2 Deutsch

In 1999, David Deutsch (Deutsch 1999) attempted to demonstrate two things within
the context of Everettian quantum mechanics: (i) the meaning of probabilistic claims
in quantum mechanics, which otherwise are ill-defined, and (ii) the fact that the
Born Rule need not be considered an extra postulate in the theory. In a more recent
paper, Deutsch stresses that (iii) his 1999 derivation of the Born Rule also overcomes
what is commonly called the “incoherence problem” in the Everettian account of
probabilities.50
As for (i), Deutsch interpreted probability as that concept which appears in
a Savage-style representation theorem within the application of rational decision
theory. In the case of quantum mechanics, Deutsch exploited the fragment of
decision theory expunged of any probabilistic notions, applying it to quantum games
in an Everettian universe – or preferences over the future of quantum mechanical
measurement outcomes associated with certain payoffs. The emergent notion of
probability is agent-dependent, in the sense that, as with Feynman’s notion of
probability (recall Sect. 7.3.3 above), probabilities arise from the actions of ideally
rational agents and have no independent existence in the universe – a branching
universe whose fundamental dynamics is deterministic.
. . . the usual probabilistic terminology of quantum theory is justifiable in light of result
of this paper, provided one understands it all as referring ultimately to the behaviour of

49 Wallace (2012), pp. 147, 148. See also Wallace (2010).

50 This is the problem of understanding how probabilities can come about when everything that can

happen does happen. See Wallace (2012, pp. 40, 41).

182 H. R. Brown and G. Ben Porath

rational decision makers. . . . [probabilistic predictions] become implications of a purely

factual theory, rather than axioms whose physical meanings are undefined.51

What distinguishes the Deutsch argument from the usual decision-theoretic

representation theorem in analogous single-world scenarios is that rational agents
are constrained to behave not just in conformity with probability theory, but the
values of the probabilities are uniquely defined by the state of the system. It is
essential in the 1999 derivation that preferences are based on states which are
represented solely by the standard pure and mixed states in quantum mechanics;
hidden variables are ruled out.
Deutsch stressed that the pivotal result concerns the special case where the
pure state is a symmetric superposition of two eigenstates of the observable being
measured (i.e. the coefficients, or amplitudes, in the superposition are equal). Based
on several rules of rationality, Deutsch showed that in this case a rational decision
maker behaves as if she believed that each of the two possible outcomes has equal
probability, and that she was maximising the probabilistic expectation value of
the payoff (expected utility). The equal probability conclusion in this case might
be considered a simple consequence of the principle of indifference, but Deutsch
is intent on showing by way of decision theory that it makes sense to assign
preferences even in the case of “indeterminate” outcomes, i.e., Everettian branching
(see point (iii) above).

7.4.3 Wallace

Wallace’s version of the Deutsch-Wallace (DW) theorem evolved through a series

of iterations, starting in 2003 and culminating in his magisterial book on the Everett
interpretation of 2012. Wallace attempted to turn Deutsch’s “minimalist” proof of
the Born Rule into a highly rigorous decision-theoretic derivation based on weaker,
but more numerous assumptions. Unlike Deutsch, Wallace adheres to a dualist
interpretation of probability, involving subjective credences and objective chances,
and sees both playing a role in the DW-theorem. We spell this out in more detail in
Sect. 7.5.1 below.
What concerns us at this point is the question raised in Sect. 7.2 whether non-
contextualism need be axiomatic in the Everettian picture. In the case of Deutsch’s
1999 proof, it is a consequence of (but not equivalent to) an implicit assumption
which Wallace was to identify and call measurement neutrality; Wallace made
it an explicit assumption in his 2003 (Wallace 2003) reworking of the Deutsch
argument. It would be generous to say that a proof of non-contextualism obtains

51 Deutsch (1999), p. 14. As Hemmo and Pitowsky noted, Deutsch’s proof, if successful, would
give “strong support to the subjective approaches to probability in general.” (Hemmo and Pitowsky
2007, p. 340).
7 Everettian Probabilities, The Deutsch-Wallace Theorem and the Principal Principle 183

in either account.52 The situation is more complicated in Wallace’s 2012 proof of

the theorem.
In his book, Wallace is at pains to show the role of what he calls non-
contextualism. His non-contextual inference theorem is the decision theoretic
analogue of Gleason’s theorem:
This theorem states that any solution to a quantum decision problem, provided that the
problem is richly structured and satisfies the assumptions of Chapter 5 and that the solution
satisfies certain rationality constraints similar to those discussed in Chapter 5, is represented
by a density operator iff it is noncontextual. (Wallace 2012, p. 214)

This is followed by proof of the claim that any solution to a quantum decision prob-
lem which is compatible with a state dependent solution must be non-contextual.53
But it is important to note how Wallace defines this condition. Informally,
an agent’s preferences conform to a probability rule that is non-contextualist
in Wallace’s terms if it assigns the same probabilities to the outcomes of a
measurement of operator X whether or not a compatible operator Y is measured at
the same time.54 After giving a more formal decision-theoretic definition, Wallace
explicitly admits that this is not exactly the principle used in Gleason’s theorem, “but
it embodies essentially the same idea”.55 We disagree. Wallace’s principle, which
for purposes of subsequent discussion we call weak non-contextualism, involves a
single measurement procedure; its violation means that a rational agent prefers “a
given act to the same (knowably the same act, in fact) under a different descrip-
tion, which violates state supervenience (and, I hope, is obviously irrational).”56
But the Gleason-related principle, or strong non-contextualism, involves mutually
incompatible procedures. Now Wallace appears to be referring to this strong non-
contextualism when he writes
It is fair to note, though, that just as a non-primitive approach to measurement allows one
and the same physical process to count as multiple abstractly construed measurements,
it also allows one and the same abstractly construed measurement to be performed by
multiple physical processes. It is then a nontrivial fact, and in a sense a physical analogue
of noncontextuality, that rational agents are indifferent to which particular process realizes
a given measurement.57

But why should rational agents be so indifferent? Because according to Wallace, it

is a consequence of the condition of measurement neutrality, which, while having
axiomatic status in both Deutsch’s 1999 version and Wallace’s 2003 version of the
DW-theorem, is a trivial corollary of the 2012 Born Rule theorem, which, to repeat,
is based on weaker assumptions. It is therefore rationally required. And

52 Similar qualms were voiced by Hemmo and Pitowsky (Hemmo and Pitowsky 2007).
53 Wallace (2012), p. 196.
54 Op. cit. p. 196.
55 Op. cit. p. 214.
56 Op. cit. p. 197.
57 Ibid.
184 H. R. Brown and G. Ben Porath

The short answer as to why is that two acts which correspond to the same abstractly
construed measurement can be transformed into the same act via processes to which rational
agents are indifferent.58

Now it seems to us, as it did to Pitowsky (see Sect. 7.2), that to contemplate
a contextual assignment of probabilities in quantum mechanics is prima facie far
from irrational, given the non-commutative property of operators associated with
measurements.59 In our view, the most plausible justification of non-contextualism
in the context of the DW theorem was given by Timpson (Timpson 2011; Sect. 5.1).
It is based on consideration of the details of the dynamics of the measurement
process in unitary quantum mechanics, and shows that nothing in the standard
account of the process supports the possibility of contextualism. However, this
argument presupposes that the standard account is independent of the Born Rule,
a supposition which deserves attention. At any rate, it should not be overlooked that
the DW-theorem applies to systems with Hilbert spaces of arbitrary dimensions,
which is a significant advantage over the proof of the Born Rule for credences using
Gleason’s theorem.

7.5 A Quantum Justification of the Principal Principle?

7.5.1 Wallace and Saunders

(i) For David Wallace and Simon Saunders, it is of great significance that Everettian
quantum mechanics (EQM) provides an underpinning, unprecedented in single-
world physics, for the Principal Principle.60 For Wallace,
. . . if an Everett-specific derivation of the Principal Principle can be given, then the Everett
interpretation solves an important philosophical problem which could not be solved under
the assumption that we do not live in a branching universe.61

For Saunders, “. . . nothing comparable has been achieved for any other physical
theory of chance.”62

58 Ibid. Note that a different, recent approach to proving that credences should be non-contextual in

quantum mechanics is urged by Steeger (Steeger, J., ’The Sufficient Coherence of Quantum States’
[unpublished manuscript]).
59 It was mentioned in Sect. 7.2 above that contextual probabilities may lead to the possibility of

superluminal signalling. But this does not imply that contextualism is irrational. Indeed, violation
of no-signalling is bound to happen in some “deviant” branches in the Everett multiverse; see
Sect. 7.6 below.
60 Perhaps this should be understood as an ‘emancipated’ PP, given that Lewis himself did not

believe in the existence of chances in a deterministic universe.

61 Wallace (2012), Footnote 26, pp. 150–151.
62 Saunders (2010), p. 184.
7 Everettian Probabilities, The Deutsch-Wallace Theorem and the Principal Principle 185

And according to Wallace, his latest approach (see below) succeeds where
Deutsch’s 1999 fails in providing a vindication of the PP.
Deutsch’s theorem . . . amounts simply to a proof of the decision theoretic link between
objective probabilities and action.63

Clearly, Wallace has a very different reading of the Deutsch theorem to ours; we see
no reference to objective probabilities therein. But Wallace is attempting to make
a substantive point: that the PP in the context of classical decision theory is not
equivalent to Papineau’s decision-theoretic link between chances and credences.
Here, briefly, is the reason.
The utility function U in the Deutsch theorem is, Wallace argues, not obviously
the same as the utility function V in classical decision theory, where the agent’s
degrees of belief may refer to unknown information. U is a feature of what Wallace
calls the Minimal decision-theoretic link between objective probability and action,
which supposedly encompasses the Deutsch theorem and Wallace’s Born Rule
theorem (though see below). In contrast, the standard classical decision theory in
which V is defined makes no mention of objective probabilities, and allows for bets
when the physical state of the system is unknown, and even when there is uncertainty
concerning the truth of the relevant physical theory.
For Wallace, the PP will be upheld in Everettian quantum mechanics only
when U and V can be shown to be equivalent (up to a harmless positive affine
transformation). If the quantum credences in the Minimal decision-theoretic link are
just subjective, or personal probabilities, then the matter is immediately resolved.
But in his 2012 book, Wallace feels the need to provide a “more direct” proof
involving a thought experiment and some further mathematics, resulting in what
he calls the utility equivalence lemma.64 Such considerations are a testament to
Wallace’s exceptional rigour and attention to detail, but for the skeptic about chances
and hence the PP (ourselves included) they seem like a considerable amount of work
for no real gain.
Let’s take another look at the Minimal decision theoretic link. The argument is,
again, that if preferences over quantum games satisfy certain plausible constraints,
the credences defined in the corresponding representation theorem are in turn
constrained to agree with branch weights. The arrow of inference goes from suitably
constrained credences to (real) numbers extracted from the (complex) state of the
system. This is somewhat out of kilt with Papineau’s decision-theoretic link, which
involves an inference from knowledge of objective probabilities to credences. And
this discrepancy is entirely innocuous, we claim, in Deutsch’s own understanding
of his 1999 theorem, where branch weights are not interpreted as objective, agent-
independent chances.
(ii) As far as we can tell, the grip that the PP has on Wallace’s thinking can be
traced back to his conviction that probabilities in physics, both in classical statistical

63 Wallace (2012), p. 237.

64 Wallace (2012), pp. 208–210 and Appendix D.
186 H. R. Brown and G. Ben Porath

mechanics and Everettian quantum mechanics (EQM), represent objective elements

of reality, despite the underlying dynamics being fundamentally deterministic. For a
philosophical realist to deny the existence of such elements of reality would thus be
tantamount to self-refutation (see Sect. 7.3.1 above). More specifically, a subjective
notion of probability, wrote Wallace in 2002,
. . . seems incompatible with the highly objective status played by probability in science in
general, and physics in particular. Whilst it is coherent to advocate the abandonment of
objective probabilities, it seems implausible: it commits one to believing, for instance, that
the predicted decay rate of radioisotopes is purely a matter of belief.65

But subjectivists do not claim that the half-life of 14 C, for example, is purely a matter
of belief. It is belief highly constrained by empirical facts. The half-life is arrived
at through Bayesian updating based on the results of ever more accurate/plentiful
statistical measurements of decay, as we saw in Sect. 7.3.3.66
For Wallace, in the case of quantum mechanics (and hence of classical statistical
mechanics, which he correctly sees as the classical limit of quantum mechanics)
the probabilistic elements of reality – chances – are (relative) branch weights, or
mod-squared amplitudes. Now no one who is a realist about the quantum state
would question whether amplitudes are agent-independent and supervenient on the
ontic universal wavefunction. But are they intrinsically “chances” of the kind that
defenders of the PP would recognise?
This is a hard question to answer, in part because the notion of chance in the
literature is so elusive. Wallace and Saunders adopt the approach of “cautious
functionalism”. Essentially, this means that branch weights act as if they were
chances, according to the PP. Here is Wallace:
. . . the Principal Principle can be used to provide what philosophers call a functional
definition of objective probability: it defines objective probability to be whatever thing fits
the ‘objective probability’ slot in the Principal Principle.67

In Saunders’ words:
[Wallace] shows that branching structures and the squared moduli of the amplitudes, in so
far as they are known, ought to play the same decision theory role [as in Papineau’s decision
theoretic link] that chances play, in so far as they are known, in one-world theories.68

And again:
The [DW] theorem demonstrates that the particular role ordinarily but mysteriously
played by physical probabilities, whatever they are, in our rational lives, is played in a
wholly perspicuous and entirely unmysterious way by branch weights and branching. It is

65 Wallace (2002), Sect. 2.7.

66 Inhis 2012 book, Wallace does acknowledge this point; see Wallace (2012, p. 138).
67 Wallace (2012), p. 141.
68 Saunders (2010), p. 184.
7 Everettian Probabilities, The Deutsch-Wallace Theorem and the Principal Principle 187

establishing that this role is played by the branch weights, and establishing that they play
all the other chance roles, that qualifies these quantities as probabilities.69

Saunders calls this reasoning a “derivation” of the PP,70 but in our view it
amounts to presupposing the PP and showing that the elusive notion of chance can
be cashed out in EQM in terms of branch weights. It hardly seems consistent with
Wallace’s express hope
. . . to find some alternative characterization of objective probability, independent of the
Principal Principle, and then prove that the Principal Principle is true for that alternatively
characterized notion.71

Note, however, that Wallace himself stresses that his Born Rule theorem is insuffi-
cient for deriving the PP, since it is “silent on what to do when the quantum state is
not known, or indeed when the agent is uncertain about whether quantum mechanics
is true.” This leads Wallace to think that the theorem, like Deutsch’s, merely
establishes the Minimal decision-theoretic link between objective probability and
action, as we have seen. As a consequence Wallace developed a decision-theoretic
“unified approach” to probabilities in EQM,72 which avoids the awkwardness of
having the concepts of probability and utility derived twice, from his Born Rule
theorem, and from appeal to something like the 2010 Greaves-Myrvold de Finetti-
inspired solution to what they call the “evidential problem” in EQM, although a
defence of the PP can still be gained by appeal (see above) to the utility equivalence
lemma.73
In our view, any such defence relies on the claim that somewhere in the
decision-theoretic reasoning an “agent-independent notion of objective probabil-
ity”74 emerges. We turn to this issue now.
(iii) In his 2010 treatment of EQM, Saunders gives an account of “what
probabilities actually are (branching structures)”.75 It is important to recognise the
role of the distinction between branching structures and amplitudes in this analysis.
The former are ‘emergent’ and non-fundamental; the latter are fundamental and
provide (in terms of their modulus squared) the numerical values of the chances:
Just like other examples of reduction, . . . [probability] can no longer be viewed as
fundamental. It can only have the status of the branching structure itself; it is ‘emergent’ . . . .
Chance, like quasiclassicality, is then an ‘effective’ concept, its meaning at the microscopic
level entirely derivative on the establishment of correlations, natural or man-made, with
macroscopic branching. That doesn’t mean that amplitudes in general . . . have no place in

69 See Saunders (2020); this paper also contains a rebuttal of the claim that the process of
decoherence, so essential to the meaning of branches in EQM, itself depends on probability
assumptions.
70 Saunders (2010).
71 Wallace (2012), p. 144.
72 Op. cit. Chap. 6.
73 Op. cit. p. 234.
74 Op. cit. p. 229.
75 Ibid.
188 H. R. Brown and G. Ben Porath

the foundations of EQM – on the contrary, they are part of the fundamental ontology – but
their link to probability is indirect. It is simply a mistake, if this reduction is successful, to
see quantum theory as at bottom a theory of probability.76

Now Everettian subjectivists (about probabilities) concur completely with this

conclusion, given that rational agents themselves have the emergent character
described in this passage. At any rate, in Saunder’s picture, at the microscopic
level, amplitudes have nothing to do with chances. Amplitudes only become
connected with chances when the appropriate experimental device correlates them
with branching structure,77 and only then, in our view, by way of the PP.
This notion of chance in EQM is weaker than that often adopted by advocates of
objective probability; indeed Wallace himself states that the functional definition of
chance has a different character from that of charge, mass or length.78 The decision-
theoretic approach to the Born Rule starts with rational agents and their preferences
in relation to quantum games; without the credences emerging from the rationality
and structure axioms, application of the PP to infer chances (branch weights) would
be impossible. In what sense then are such chances “out there”, independent of the
rational cogitations of agents?
Wallace’s answer to this question appears in the Second Interlude of his 2012
book, and in which the Author replies to objections or queries raised by the Sceptic.
Here is the relevant passage:
SCEPTIC: Do you really find it acceptable to regard quantum probabilities as defined via
decision theory? Shouldn’t things like the decay rate of tritium be objective, rather than
defined via how much we’re prepared to bet on them?
AUTHOR: The quantum probabilities are objective. In fact, it’s clearer in the Everettian
context what those objective things are: they’re relative branch weights. They’d be
unchanged even if there were no bets being made.
SCEPTIC: In that case, what’s the point of the decision theory?
AUTHOR: Although branch weights are objective, what makes it true that branch
weight=probability is the way in which branch weights figure in the actions of (ideally)
rational agents. . . . 79

It is easy to see the tension between the Author’s replies if one thinks of the
objective probabilities emerging from the DW-theorem as having an existence
independent of betting agents.80 But such is not what the application of the PP yields
here. And note again that the branch weights yielded by the wavefunction are not
objective probabilities (according to the argument) until the branches are effectively

76 Saunders (2010), p. 182.

77 Simon Saunders, private communication.
78 Wallace (2012), p. 144. Compare this with the views expressed in Sect. 2.1 above.
79 Wallace 2012, p. 249.
80 Wallace, ibid, compares branch weights with physical money, in our view a very apt analogy.

A dollar note, e.g., has no more intrinsic status as money than branch weights have the status of
probability, until humans confer this status on it. This point is nicely brought out by Harari, who
writes that the development of money “involved the creation of a new inter-subjective reality that
exists solely in people’s shared imagination.” See Harari 2014, chapter 10.
7 Everettian Probabilities, The Deutsch-Wallace Theorem and the Principal Principle 189

defined by the process of decoherence. In an important sense, adoption of something

like this derivative notion of chance in the Saunders-Wallace approach is inevitable
for chance “realists” who defend EQM. If agent-independent chances existed, then
there would be more to Everettian ontology than just the universal wavefunction,
contrary to the aims of the approach.81
The subjectivist, on the other hand, who views the quantum probabilities
as credences, may take (subject to doubts raised in the next section) the DW-
theorem to show, remarkably, that there are objective features of the Everettian
“multiverse” that together with rationality principles constrain credences (which are
not fundamental elements in EQM) to align with the Born Rule. There is arguably no
strict analogue in classical, one-world physics, as Deutsch emphasised. But to call
relative branch weights “objective probabilities” is a mere façon de parler, and the
temptation to functionally define them as such only reflects an a priori commitment
of questionable merit to the validity of the PP.
It is noteworthy that in his 2010 treatment of the DW-theorem, Wallace says the
following:
The decision-theoretic language in which this paper is written is no doubt necessary to
make a properly rigorous case and to respond to those who doubt the very coherence of
Everettian probability, but in a way the central core of the argument is not decision-theoretic
at all. What is really going on is that the quantum state has certain symmetries and the
probabilities are being constrained by those symmetries.82

Note that the probabilities in the argument are credences; the exercise is at heart
the application of a sophisticated variant of the principle of indifference based on
symmetry. This makes the nature of quantum probabilities essentially something
Laplace would recognise, but with the striking new feature mentioned at the
beginning of this section. To reify the resulting quantitative values of the credences
(branch weights) as chances seems to us both unnecessary and ill-advised; it would
be like telling Laplace that his credence in the outcome ‘heads’ is a result (in part)
of an objective probability inherent in the individual (unbiased) coin.83

81 Wallace describes the Everettian program as interpreting the “bare quantum formalism” – which
itself makes no reference to probability (Wallace 2012, p. 16) – in a “straightforwardly realist way”
without modifying quantum mechanics (op. cit., p. 36).
82 Wallace (2010), pp. 259, 260. Saunders has recently also stressed the role of symmetry in the

DW theorem; see Saunders (2020, Sect. 7.6).

83 Our approach to probability in EQM is very similar to the “pragmatic” view of Bacciagaluppi

and Ismael, in their thoughtful 2015 review of Wallace’s 2012 book (of which more in Sect. 7.6
below):
According to such a view, the ontological content of the theory makes no use of
probabilities. There is a story that relates the ontology to the evolution of observables
along a family of decoherent histories, and probability is something that plays a role in
the cognitive life of an agent whose experience is confined to sampling observables along
such a history. In so doing, one would still be doing EQM . . . (Bacciagaluppi and Ismael
2015, Sect. 3.2).
190 H. R. Brown and G. Ben Porath

7.5.2 Earman

Recently, John Earman (2018) has also argued that something like the PP is
a “theorem” of non-relativistic quantum mechanics, but now in a single-world
context. Briefly, the argument goes like this.
Earman starts the argument within the abstract algebraic formulation of the
theory. A “normal” quantum state ω is defined as a normed positive linear functional
on the von Neumann algebra of bounded observables B(H) operating on a separable
Hilbert space H, with a density operator representation. This means that there is a
trace class operator ρ on H with T r(ρ) = 1, such that ω(A) = T r(ρA) for all A ∈
B(H). Normal quantum states induce quantum probability functions P r ω on the
lattice of projections P(B(H)): P r ω (E) = ω(E) = T r(ρE) for all E ∈ P(B(H)).
Earman takes the quantum state to be an objective feature of the physical system,
and infers that the probabilities induced by them are objective.84 This is spelt out
by showing how pure state preparation can be understood non-probabilistically
using the von Neumann projection postulate in the case of yes-no experiments, and
using the law of large numbers to relate the subsequent probabilities induced by the
prepared state to frequencies in repeated measurements. All this Earman calls the
“top-down” approach to objective quantum probabilities.
Credences are now introduced into the argument by considering a “bottom-
up” approach’ in which probability functions on P(B(H)) are construed as the
credence functions of actual or potential Bayesian agents. States are construed as
bookkeeping devices used to keep track of credence functions. By considering
again a (non-probabilistic) state preparation procedure, and the familiar Lüders
updating rule, Earman argues that a rational agent will in this case, in the light
of Gleason’s theorem, adopt a credence function on P(B(H)) which is precisely the
objective probability induced by the prepared pure state (as defined in the previous
paragraph).85 Thus, Earman is intent on showing
that there is a straightforward sense in which no new principle of rationality is needed to
bring rational credences over quantum events into line with the events’ objective chances –
the alignment is guaranteed by as a theorem of quantum probability, assuming the credences
satisfy a suitable form of additivity. (Earman 2018)

Here are a few remarks on Earman’s argument.

1. Unless the PP in its original guise is interpreted as analytically true, or providing
an implicit definition of chance, or, following Myrvold, it is not about real

However, we do not follow these authors in their attempt to define chances consonant with such
probabilities and the PP. (We are grateful to David Wallace for drawing to our attention this review
paper.)
84 Op. Cit. p. 16.
85 Recall that in the typical case of the infinite dimensional, separable Hilbert space H, a condition

of Gleason’s theorem is that the probability function must be countably additive; see Sect. 7.2
above.
7 Everettian Probabilities, The Deutsch-Wallace Theorem and the Principal Principle 191

chances at all, the principle is virtually by definition a principle or rule of

rationality.86 Were it valid, the appeal to Gleason’s theorem in Earman’s account
would be redundant: credences would track the chances supposedly embedded
in the algebraic formalism. Earman purports to show that chances and credences
in quantum mechanics have separate underpinnings, and that their numerical
values coincide when referring to the same measurement outcomes. Rather than
providing a quantum theoretical derivation of the PP, Earman is essentially
rejecting the PP and attempting to prove a connection between chances and
credences whose nature is not that of an extra principle of rationality.
2. The argument in the top-down view is, however, restricted to the algebraic
formulation of quantum mechanics that Earman advocates. Other approaches,
such as the early de Broglie-Bohm theory and the Everett picture, also take the
quantum state – or that part of it related to the Hilbert space – to represent
a feature of reality. But it is essentially defined therein as a solution to the
time (in-)dependent Schödinger equation(s), with no a priori connection between
the state and probabilities of measurement outcomes, unless in the case of
the Everett picture the decisions of agents are taken into account.87 In the
algebraic approach, this connection is baked in, at a very heavy cost: the terms
“measurement” and “objective probability” are primitive.
3. Earman’s argument is subject to the same criticism that was raised above in
relation to Pitowsky’s derivation of the Born Rule: to the extent that they both rely
on Gleason’s theorem, credence functions are assumed to be not just countably
additive but non-contextual. And the derivation breaks down for Hilbert spaces
with dimensions less than 3, as Earman recognises.

7.6 The Refutability Issue for Everettian Probability

There is an important aspect of probability as a scientific notion that has been

overlooked so far in this paper. It is the falsifiable status of probabilistic claims,
as stressed by Karl Popper and more recently Donald Gillies.88 Deutsch himself
gives it prominent status in his 2016 paper:
If a theory attaches numbers pi to possible results ai of an experiment, and calls those
numbers ‘probabilities’, and if, in one or more instances of the experiment, the observed
frequencies of the ai differ significantly, according to some [preordained] statistical test,
from the pi , then a scientific problem should be deemed to exist. (Deutsch 2016)

In relation to dynamical collapse theories and de Broglie-Bohm pilot wave theory,

Deutsch argues that without this methodological rule the probabilities therein are
merely decorative, and with it the theories are problematic. The rule, says Deutsch,

86 See, for instance, (Wallace 2012, section 4.10).

87 But recall the complication referred to in footnote 64 above.
88 See Gillies (2000, chapter 7).
192 H. R. Brown and G. Ben Porath

“is not a supposed law of nature, nor is it any factual claim about what happens
in nature (the explicanda), nor is it derived from one. . . . And one cannot make an
explanation problematic merely by declaring it so.” p. 29.
For Deutsch, it is a triumph of the DW-theorem that in Everettian quantum
mechanics this problem – that the methodological falsifiability rule is mere whim –
is avoided. The theorem shows that
. . . rational gamblers who knew Everettian quantum theory, . . . and have . . . made no
probabilistic assumptions, when playing games in which randomisers were replaced by
quantum measurements, would place their bets as if those were randomisers, i.e. using the
[Born] probabilistic rule . . . according to the methodological [falsifiability] rule [above].
p. 31.

Furthermore, it is argued that “the experimenter – who is now aware of the same
evidence and theories as they are – must agree with them” when one of two
gamblers regards his theory as having been refuted. The argument is subtle, if not
convoluted, and requires appeal on the part of gamblers to a non-probabilistic notion
of expectation that Deutsch introduces early in the 2016 paper.
Perhaps a simpler approach works. Subjectivists, as much as objectivists, aim to
learn from experience; despite the immutability of their conditional probabilities
(see Sect. 7.3.3) they of course update their probabilities conditional on new
information. As we saw, this is the basis of their version of the law of large numbers.
In fact, it might be considered part of the scientific method for subjectivists, or of
the very meaning of probabilities in science, that updating is deemed necessary in
cases where something like the Popper-Gillies-Deutsch falsifiability rule renders the
theory in question problematic. In practice, such behaviour is indistinguishable from
that of objectivists about probability in cases of falsification (even in the tentative
sense advocated by Deutsch and others). But this line of reasoning does not make
special appeal to a branching universe, and thus is not in the spirit of Deutsch’s
argument.
The reader can decide which, if either, of these approaches is right. But now an
interesting issue arises regarding the DW-theorem, as a number of commentators
have noticed.
Given the branching created in (inter alia) repeated measurement processes, it
is inevitable that in some branches statistics of measurement outcomes will be
obtained that fail to adhere to the Born Rule, whatever reasonable preordained
statistical test is chosen. These are sometimes called deviant branches. What will
Everettian-favourable observers in such branches – conclude?89 If they adhere to

89 In his 1999 paper, Deutsch assumed that agents involved in his argument were initially adherents

of the non-probabilistic part of Everettian theory. Wallace was influenced by the work of Greaves
and Myrvold (Greaves and Myrvold 2010) who developed a confirmation theory suitable for
branching as well as non-branching universes, but he went on to develop his own confirmation
theory as part of what he called a “unified approach” to probabilities in EQM (see Sect. 7.4.1(ii)
above). However, Deutsch, in his 2016 paper, follows Popperian philosophy and rejects any notion
of theory confirmation, thereby explicitly sidelining the work of Greaves and Myrvold, despite the
fact that it contains a falsifiability element as well as confirmation.
7 Everettian Probabilities, The Deutsch-Wallace Theorem and the Principal Principle 193

the Popper-Gillies-Deutsch falsifiability rule, they must conclude that their theory is
in trouble. Wallace and Saunders correctly point out that from the point of view of
the DW-theorem, they happen just to be unlucky.90 But such observers will have
no option but to question Everettian theory. To say that statistics trump theory,
including principles of rationality ultimately based on symmetries,91 is just to say
that the falsifiability rule is embraced, and not to embrace it for subjectivists like
Deutsch is to confine the probabilistic component of quantum mechanics to human
psychology.
But what part of the theory is to be questioned? Read has recently argued that
it is the non-probabilistic fragment of EQM – the theory an Everettian accepts
before adopting the Born Rule. In this case, observers in deviant branches could
question the physical assumptions involved in the DW-theorem (for example, that
the agent’s probabilities supervene solely on the wavefunction), and thus consider it
inapplicable to their circumstances.
Hemmo and Pitowsky, on the other hand, argued in 2007 (Hemmo and Pitowsky
2007) that such an observer could reasonably question the rationality axioms that
lead to non-contextualism in (either version of) the DW-theorem. Here is Wallace’s
2012 reply.
. . . it should be clear that the justification used in this chapter is not available to operational-
ists, for (stripping away the technical detail) the Everettian defence of non-contextuality is
that two processes are decision-theoretically equivalent if they have the same effect on the
physical state of the system, whatever it is. Since operationalists deny that there is a physical
state of the system, this route is closed to them.92

Now we saw in Sect. 7.2 (footnote 4) that Pitowsky, in particular, denies the physical
reality of the quantum state. But it is unclear to us whether this is relevant to the
matter at hand. Wallace seems to be referring to what we called in Sect. 7.4.3
weak non-contextualism, whereas what Hemmo and Pitowsky have in mind is
strong non-contextualism. Wallace proceeds to give a detailed classical decision
problem (analogous to the quantum decision problem) which contains a condition of
“classical noncontextuality” that again is justified if it assumed “that the state space
really is a space of physical states”. But in a classical theory, no analogue exists
of non-commutativity and the incompatibility of measurement procedures involved
in strong non-contextualism. It seems to us that the concern raised by Hemmo and
Pitowsky is immune to Wallace’s criticism; it does not depend on adopting a non-
“operationalist” stand in relation to the quantum state, and applies just as much to
non-deviant as to deviant branches.93 But we mentioned in Sect. 7.3.3 above that a
strong plausibility argument for non-contextualism in Wallace’s decision-theoretic
approach has been provided by Timpson.

90 See Wallace (2012, p. 196) and Saunders (2020). Saunders again correctly points out that in an
infinite non-branching universe an analogous situation holds.
91 Read (2018), p. 138.
92 Wallace (2012), p. 226.
93 However, Hemmo and Pitowsky also argue that
194 H. R. Brown and G. Ben Porath

In the Copenhagen interpretation of quantum mechanics, quantum bets are in

effect inferred from observed frequencies: the Born Rule has as its justification
nothing more than past statistics. This option is open to the Everettian (though
adherents of the DW-theorem hope for more). Indeed, whether they are objectivists
or subjectivists concerning probability, their reasoning here would be no different
from that described in Sect. 7.3.3 in relation to the determination of the half-life
of 14 C. And this threatens now to make the DW-theorem redundant in non-deviant
branches.94 After all, if statistics trump theory (see above) this should be the case as
much for agents in non-deviant as in deviant branches.
Read, following Richard Dawid and Karim Thébault (Dawid and Thébault 2014),
has essentially endorsed the redundancy – or what they call the irrelevancy –
conclusion, with one caveat.
. . . the central case in which DW could have any relevance is the . . . scenario in which EQM
is believed, but the agent in question has no statistical evidence for or against the theory.
DW might deliver a tighter link between subjective probabilities and branch weights, and
therefore put the justification of PP, and the status of quantum mechanical branch weights
as objective probabilities, on firmer footing. However, DW is not necessary to establish the
rationality of betting in accordance with Born rule probabilities tout court.95

While we have doubts above as to whether the DW-theorem does justify the PP,
Read is obviously right that the past-statistics-driven route to the Born Rule is not
available to an agent ignorant of the statistics! But such an epistemologically-limited
agent would be very hard to find in practice.

. . . in the many worlds theory, not only Born’s rule but any probability rule is meaningless.
The only way to solve the problem . . . is by adding to the many worlds theory some
stochastic element. (Hemmo and Pitowsky 2007, p. 334)
This follows from the claim that the shared reality of all of the multiple branches in EQM entails
“one cannot claim that the quantum probabilities might be inferred (say, as an empirical conjecture)
from the observed frequencies.” (p. 337). This is a variant of the “incoherence” objection to
probabilities in EQM (see footnote 45 above), but in our opinion it is not obviously incoherent
to bet on quantum games in a branching scenario (see Wallace 2010).
94 Although in their above-mentioned review, Bacciagaluppi and Ismael do not go quite this

far, they do not regard the DW-theorem as necessary for “establishing the intelligibility of
probabilities” in EQM. By appealing to standard inductive practice based on past frequencies,
they argue:
The inputs [frequencies] are solidly objective, but the outputs [probabilities] need not be
reified.. . . In either the classical or the Everettian setting, probability emerges uniformly as
a secondary or derived notion, and we can tell an intelligible story about the role it plays in
setting credences. (Bacciagaluppi and Ismael 2015, section 3.2).

95 Read (2018), p. 140. Note that Read, like Dawid and Thébault, rests the rationality of the
statistics-driven route on the Bayesian confirmation analysis given by Greaves and Myrvold op.
cit..
7 Everettian Probabilities, The Deutsch-Wallace Theorem and the Principal Principle 195

7.7 Conclusions

The principal motivation for this paper was the elucidation of the difference between
the views of David Deutsch and David Wallace as regards the significance of the
DW-theorem. Deutsch interprets the probabilities therein solely within the Ramsay-
de Finetti-Savage tradition of rational choice, now in the context of quantum games.
Wallace defends a dualistic interpretation involving objective probabilities (chances)
and credences, and, like Simon Saunders, argues that his more rigorous version
of the theorem provides, at long last, a “derivation” or “proof” of David Lewis’
Principal Principle.
We have questioned whether this derivation of the PP is sound. In our view, the
Wallace-Saunders argument presupposes the PP and provides within it a number
to fill the notorious slot representing chances. For proponents of reified chances in
the world, this is, nonetheless, a remarkable result. But it raises awkward questions
about the scope of the ontology in EQM, and for subjectivists about probability like
Deutsch the concerns with the PP should seem misguided in the first place. Our
own inclinations are closer to those of Deutsch, and in key respects mirror those of
Bacciagaluppi and Ismael in their discussion of the DW-theorem.
We have also questioned the doubts expressed by Hemmo and Pitowsky as to
whether all the rationality assumptions in the DW-theorem are compelling, with
particular emphasis on the role of non-contextualism. Finally, by considering the
falsifiable nature of probabilities in science, we regard the complaint by Dawid and
Thébault, and largely endorsed by Read, that the DW-theorem is, for all practical
purposes, redundant, a serious challenge to those who endorse the arguments of the
authors of the theorem.

Acknowledgements The authors are grateful to the editors for the invitation to contribute to this
volume. HRB would like to acknowledge the support of the Notre Dame Institute for Advanced
Study; this project was started while he was a Residential Fellow during the spring semester
of 2018. Stimulating discussions over many years with Simon Saunders and David Wallace are
acknowledged. We also thank David Deutsch, Chiara Marletto, James Read and Christopher
Timpson for discussions, and Guido Bacciagaluppi, Donald Gillies and Simon Saunders for helpful
critical comments on the first draft of this paper. We dedicate this article to the memory of Itamar
Pitowsky, who over many years was known to and admired by HRB, and to Donald Gillies, teacher
to HRB and guide to the philosophy of probability.

References

Albrecht, A., & Phillips D. (2014). Origin of probabilities and their application to the multiverse.
Physical Review D, 90, 123514.
Bacciagaluppi, G., & Ismael, J. (2015). Essay review: The emergent multiverse. Philosophy of
Science, 82, 1–20.
Bacciagaluppi, G. (2020). Unscrambling subjective and epistemic probabilities. This volume,
http://philsci-archive.pitt.edu/16393/1/probability%20paper%20v4.pdf
196 H. R. Brown and G. Ben Porath

Barnum, H. (2003). No-signalling-based version of Zurek’s derivation of quantum probabilities: A

note on “environment-assisted invariance, entanglement, and probabilities in quantum physics”.
quant-ph/0312150.
Bé, M. M., & Chechev, V. P. (2012). 14 C – Comments on evaluation of decay data. www.nucleide.
org/DDEP_WG/Nucleides/C-14_com.pdf. Accessed 01 June 2019.
Bell, J. S. (1966). On the problem of hidden variables in quantum mechanics. Reviews of Modern
Physics, 38, 447–452.
Bricmont, J. (2001). Bayes, Boltzmann and Bohm: Probabilities in physics. In J. Bricmont, D.
Dürr, M. C. Galavotti, G. Ghirardi, F. Petruccione, & N. Zangi (Eds.), Chance in physics.
Foundations and perspectives. Berlin/Heidelberg/New York: Springer.
Brown, H. R. (2011). Curious and sublime: The connection between uncertainty and probability in
physics. Philosophical Transactions of the Royal Society, 369, 1–15. https://doi.org/10.1098/
rsta.2011.0075
Brown, H. R. (2017). Once and for all; The curious role of probability in the Past Hypothesis.
http://philsci-archive.pitt.edu/id/eprint/13008. Forthcoming In D. Bedingham, O. Maroney, &
C. Timpson (Eds.). (2020). The quantum foundations of statistical mechanics. Oxford: Oxford
University Press.
Brown, H. R. (2019). The reality of the wavefunction: Old arguments and new. In A. Cordero (Ed.),
Philosophers look at quantum mechanics (Synthese Library 406, pp. 63–86). Springer. http://
philsci-archive.pitt.edu/id/eprint/12978
Brown, H. R., & Svetlichny, G. (1990). Nonlocality and Gleason’s Lemma. Part I. Deterministic
theories. Foundations of Physics, 20, 1379–1387.
Busch, P. (2003). Quantum states and generalized observables: A simple proof of Gleasons
theorem. Physical Review Letters, 91, 120403.
Dawid, R., & Thébault, K. (2014). Against the empirical viability of the Deutsch-Wallace-Everett
approach to quantum mechanics. Studies in History and Philosophy of Modern Physics, 47,
55–61.
de Finetti, B. (1964). Foresight: Its logical laws, its subjective sources. English translation. In: J.
E. Kyburg & H. E. Smokler (Eds.), Studies in subjective probability (pp. 93–158). New York:
Wiley.
Deutsch, D. (1999). Quantum theory of probability and decisions. Proceedings of the Royal Society
of London, A455, 3129–3137.
Deutsch, D. (2016). The logic of experimental tests, particularly of Everettian quantum theory.
Studies in History and Philosophy of Modern Physics, 55, 24–33.
Earman, J. (2018) The relation between credence and chance: Lewis’ “Principal Principle” is a
theorem of quantum probability theory. http://philsci-archive.pitt.edu/14822/
Feynman, R. P., Leighton, R. B., & Matthew Sands, M. (1965). The Feynman lectures on physics
(Vol. 1). Reading: Addison-Wesley.
Gillies, D. (1972). The subjective theory of probability. British Journal for the Philosophy of
Science, 23(2), 138–157.
Gillies, D. (1973). An objective theory of probability. London: Methuen & Co.
Gillies, D. (2000). Philosophical theories of probability. London: Routledge.
Galavotti, M. C. (1989). Anti-realism in the philosophy of probability: Bruno de Finetti’s
Subjectivism. Erkenntnis, 31, 239–261.
Gleason, A. (1957). Measures on the closed subspaces of a Hilbert space. Journal of Mathematics
and Mechanics, 6, 885–894.
Greaves, H., & Myrvold, W. (2010). Everett and evidence. In S. Saunders, J. Barrett, A. Kent, &
D. Wallace (Eds.), Many worlds? Everett, quantum theory, and reality (pp. 264–304). Oxford:
Oxford University Press. http://philsci-archive.pitt.edu/archive/0004222
Hacking, I. (1984). The emergence of probability. Cambridge: Cambridge University Press.
Hall, N. (2004). Two mistakes about credence and chance. Australasian Journal of Philosophy, 82,
93–111.
Handfield, T. (2012). A philosophical guide to chance. Cambridge: Cambridge University Press.
Harari, Y. N. (2014). Sapiens. A brief history of humankind. London: Harvill-Secker.
7 Everettian Probabilities, The Deutsch-Wallace Theorem and the Principal Principle 197

Hemmo, M., & Pitowsky, I. (2007). Quantum probability and many worlds. Studies in History and
Philosophy of Modern Physics, 38, 333–350.
Hoefer, C. (2019). Chance in the world: A Humean guide to objective chance. (Oxford Studies in
Philosophy of Science). Oxford, Oxford University Press, 2019.
Hume, D. (2008). An enquiry concerning human understanding. Oxford: Oxford University Press.
Ismael, J. (1996). What chances could not be. British Journal for the Philosophy of Science, 47(1),
79–91.
Ismael, J. (2019, forthcoming). On Chance (or, Why I am only Half-Humean). In S. Dasgupta
(Ed.), Current controversies in philosophy of science. London: Routledge.
Jaynes, E. T. (1963). Information theory and statistical mechanics. In: G. E. Uhlenbeck et al. (Eds.),
Statistical physics. (1962 Brandeis lectures in theoretical physics, Vol. 3, pp. 181–218). New
York: W.A. Benjamin.
Kochen, S., & Specker, E. P. (1967). The problem of hidden variables in quantum mechanics.
Journal of Mathematics and Mechanics, 17, 59–87.
Lewis, D. (1980). A subjectivist’s guide to objective chance. In: R. C. Jeffrey (Ed.), Studies in
inductive logic and probability (Vol. 2, pp. 263–293). University of California Press (1980).
Reprinted in: Lewis, D. Philosophical papers (Vol. 2, pp. 83–132). Oxford: Oxford University
Press.
Maris, J. P, Navrátil, P., Ormand, W. E., Nam, H., & Dean, D. J. (2011). Origin of the anomalous
long lifetime of 14 C. Physical Review Letters, 106, 202502.
Maudlin, T. (2007). What could be objective about probabilities? Studies in History and Philosophy
of Modern Physics, 38, 275–291.
McQueen, K. J., & Vaidman, L. (2019). In defence of the self-location uncertainty account of
probability in the many-worlds interpretation. Studies in History and Philosophy of Modern
Physics, 66, 14–23.
Myrvold, W. C. (2016). Probabilities in statistical mechanics, In C. Hitchcock & A. Hájek (Eds.),
Oxford handbook of probability and philosophy. Oxford: Oxford University Press. Available at
http://philsci-archive.pitt.edu/9957/
Page, D. N. (1994). Clock time and entropy. In J. J. Halliwell, J. Pérez-Mercader, & W. H.
Zurek (Eds.), The physical origins of time asymmetry, (pp. 287–298). Cambridge: Cambridge
University Press.
Page, D. N. (1995). Sensible quantum mechanics: Are probabilities only in the mind? arXiv:gr-
qc/950702v1.
Papineau, D. (1996). Many minds are no worse than one. British Journal for the Philosophy of
Science, 47(2), 233–241.
Pattie, R. W. Jr., et al. (2018). Measurement of the neutron lifetime using a magneto-gravitational
trap and in situ detection. Science, 360, 627–632.
Pitowsky, I. (1994). George Boole’s ‘conditions of possible experience’ and the quantum puzzle.
British Journal for the Philosophy of Science, 45(1), 95–125.
Pitowsky, I. (2006). Quantum mechanics as a theory of probability. In W. Demopoulos & I.
Pitowsky (Eds.), Physical theory and its interpretation (pp. 213–240). Springer. arXiv:quant-
phys/0510095v1.
Read, J. (2018). In defence of Everettian decision theory. Studies in History and Philosophy of
Modern Physics, 63, 136–140.
Saunders, S. (2005). What is probability? In A. Elitzur, S. Dolev, & N. Kolenda (Eds.), Quo vadis
quantum mechanics? (pp. 209–238). Berlin: Springer.
Saunders, S. (2010). Chance in the Everett interpretation. In S. Saunders, J. Barrett, A. Kent, &
D. Wallace (Eds.), Many worlds? Everett, quantum theory, and reality (pp. 181–205). Oxford:
Oxford University Press.
Saunders, S. (2020), The Everett interpretation: Probability. To appear in E. Knox & A. Wilson
(Eds.), Routledge Companion to the Philosophy of Physics. London: Routledge.
Skyrms, B. (1984). Pragmatism and empiricism. New Haven: Yale University Press.
Strevens, M. (1999). Objective probability as a guide to the world, philosophical studies. An
International Journal for Philosophy in the Analytic Tradition, 95(3), 243–275.
198 H. R. Brown and G. Ben Porath

Svetlichny, G. (1998). Quantum formalism with state-collapse and superluminal communication.

Foundations of Physics, 28(2), 131–155.
Timpson, C. (2011). Probabilities in realist views of quantum mechanics, chapter 6. In C. Beisbart
& S. Hartmann (Eds.), Probabilities in physics. Oxford: Oxford University Press.
Tipler, F. J. (2014). Quantum nonlocality does not exist. Proceedings of the National Academy of
Sciences, 111(31), 11281–11286.
Vaidman, L. (2020). Derivations of the Born Rule. This volume, http://philsci-archive.pitt.edu/
15943/1/BornRule24-4-19.pdf
Wallace, D. (2002). Quantum probability and decision theory, Revisited, arXiv:quant-
ph/0211104v1.
Wallace, D. (2003). Everettian rationality: Defending Deutsch’s approach to probability in the
Everett interpretation. Studies in History and Philosophy of Modern Physics, 34, 415–439.
Wallace, D. (2010). How to prove the born rule. In: S. Saunders, J. Barrett, A. Kent, & D. Wallace
(Eds.), Many worlds? Everett, quantum theory, and reality (pp. 227–263). Oxford: Oxford
University Press.
Wallace, D. (2012). The emergent universe. Quantum theory according to the Everett interpreta-
tion. Oxford: Oxford University Press.
Wallace, D. (2014). Probability in physics: Statistical, stochastic, quantum. In A. Wilson (Ed.),
Chance and temporal asymmetry (pp. 194–220). Oxford: Oxford University Press. http://
philsci-archive.pitt.edu/9815/1/wilson.pdf
Wright, V. J., & Weigert, S. (2019). A Gleason-type theorem for qubits based on mixtures of
projective measurements. Journal of Physics A: Mathematical and Theoretical, 52(5), 055301.
Chapter 8
‘Two Dogmas’ Redux

Jeffrey Bub

Abstract I revisit the paper ‘Two dogmas about quantum mechanics,’ co-authored
with Itamar Pitowsky, in which we outlined an information-theoretic interpretation
of quantum mechanics as an alternative to the Everett interpretation. Following the
analysis by Frauchiger and Renner of ‘encapsulated’ measurements (where a super-
observer, with unrestricted ability to measure any arbitrary observable of a complex
quantum system, measures the memory of an observer system after that system
measures the spin of a qubit), I show that the Everett interpretation leads to modal
contradictions. In this sense, the Everett interpretation is inconsistent.

Keywords Measurement problem · Information-theoretic interpretation · Everett

interpretation · Wigner’s friend · Encapsulated measurements ·
Frauchiger-Renner argument

8.1 Introduction

About ten years ago, Itamar Pitowsky and I wrote a paper, ‘Two dogmas about quan-
tum mechanics’ (Bub and Pitowsky 2010), in which we outlined an information-
theoretic interpretation of quantum mechanics as an alternative to the Everett
interpretation. Here I revisit the paper and, following Frauchiger and Renner
(2018), I show that the Everett interpretation leads to modal contradictions in
‘Wigner’s-Friend’-type scenarios that involve ‘encapsulated’ measurements, where
a super-observer (which could be a quantum automaton), with unrestricted ability
to measure any arbitrary observable of a complex quantum system, measures the

J. Bub ()
Department of Philosophy, Institute for Physical Science and Technology, Joint Center for
Quantum Information and Computer Science, University of Maryland, College Park, MD, USA
e-mail: jbub@umd.edu

© Springer Nature Switzerland AG 2020 199

M. Hemmo, O. Shenker (eds.), Quantum, Probability, Logic,
Jerusalem Studies in Philosophy and History of Science,
https://doi.org/10.1007/978-3-030-34316-3_8
200 J. Bub

memory of an observer system (also possibly a quantum automaton) after that

system measures the spin of a qubit. In this sense, the Everett interpretation is
inconsistent.

8.2 The Information-Theoretic Interpretation

The salient difference between classical and quantum mechanics is the noncom-
mutativity of the algebra of observables of a quantum system, equivalently the
non-Booleanity of the algebra of two-valued observables representing properties
(for example, the property that the energy of the system lies in a certain range of
values, with the two eigenvalues representing ‘yes’ or ‘no’), or propositions (the
proposition asserting that the value of the energy lies in this range, with the two
eigenvalues representing ‘true’ or ‘false’). The two-valued observables of a classical
system form a Boolean algebra, isomorphic to the Borel subsets of the phase space
of the system. The transition from classical to quantum mechanics replaces this
Boolean algebra by a family of ‘intertwined’ Boolean algebras, to use Gleason’s
term (Gleason 1957), one for each set of commuting two-valued observables, repre-
sented by projection operators in a Hilbert space. The intertwinement precludes the
possibility of embedding the whole collection into one inclusive Boolean algebra, so
you can’t assign truth values consistently to the propositions about observable values
in all these Boolean algebras. Putting it differently, there are Boolean algebras in the
family of Boolean algebras of a quantum system, for example the Boolean algebras
for position and momentum, or for spin components in different directions, that
don’t fit together into a single Boolean algebra, unlike the corresponding family for
a classical system. In this non-Boolean theory, probabilities are, as von Neumann
put it, ‘sui generis’ (von Neumann 1937), and ‘uniquely given from the start’ (von
Neumann 1954, p. 245) via Gleason’s theorem, or the Born rule, as a feature of
the geometry of Hilbert space, related to the angle between rays in Hilbert space
representing ‘pure’ quantum states.1
The central interpretative question for quantum mechanics as a non-Boolean
theory is how we should understand these ‘sui generis’ probabilities, since they are
not probabilities of spontaneous transitions between quantum states, nor can they be
interpreted as measures of ignorance about quantum properties associated with the
actual values of observables prior to measurement.
The information-theoretic interpretation is the proposal to take Hilbert space
as the kinematic framework for the physics of an indeterministic universe, just
as Minkowski space provides the kinematic framework for the physics of a non-

1 This conceptual picture applies to quantum mechanics on a finite-dimensional Hilbert space. A

restriction to ‘normal’ quantum states is required for quantum mechanics formulated with respect
to a general von Neumann algebra, where a generalized Gleason’s theorem holds even for quantum
probability functions that are not countably additive. Thanks to a reviewer for pointing this out.
8 ‘Two Dogmas’ Redux 201

Newtonian, relativistic universe.2 In special relativity, the geometry of Minkowski

space imposes spatio-temporal constraints on events to which the relativistic dynam-
ics is required to conform. In quantum mechanics, the non-Boolean projective
geometry of Hilbert space imposes objective kinematic (i.e., pre-dynamic) proba-
bilistic constraints on correlations between events to which a quantum dynamics
of matter and fields is required to conform. In this non-Boolean theory, new sorts
of nonlocal probabilistic correlations are possible for ‘entangled’ quantum states of
separated systems, where the correlated events are intrinsically random, not merely
apparently random like coin tosses (see Bub 2016, Chapter 4).
In (Pitowsky 2007), Pitowsky distinguished between a ‘big’ measurement prob-
lem and a ‘small’ measurement problem. The ‘big’ measurement problem is the
problem of explaining how measurements can have definite outcomes, given the
unitary dynamics of the theory. The ‘small’ measurement problem is the problem
of accounting for our familiar experience of a classical or Boolean macroworld,
given the non-Boolean character of the underlying quantum event space. The
‘big’ problem is the problem of explaining how individual measurement outcomes
come about dynamically, i.e., how something indefinite in the quantum state of a
system before measurement can become definite in a measurement. There is nothing
analogous to this sort of transition in classical physics, where transitions are always
between an initial state of affairs specified by what is and is not the case (equivalent
to a 2-valued homomorphism on a Boolean algebra) to a final state of affairs with
a different specification of what is and is not the case. The ‘small’ problem is the
problem of explaining the emergence of an effectively classical probability space
for the macro-events we observe in a measurement.
The ‘big’ problem is deflated as a pseudo-problem if we reject two ‘dogmas’
about quantum mechanics. The first dogma is the view that measurement in a
fundamental mechanical theory should be treated as a dynamical process, so
it should be possible, in principle, to give a complete dynamical analysis of
how individual measurement outcomes come about. The second dogma is the
interpretation of a quantum state as a representation of physical reality, i.e., as the
‘truthmaker’ for propositions about the occurrence and non-occurrence of events,
analogous to the ontological significance of a classical state.
The first dogma about measurement is an entirely reasonable demand for a
fundamental theory of motion like classical mechanics. But noncomutativity or
non-Booleanity makes quantum mechanics quite unlike any theory we have dealt
with before in the history of physics, and there is no reason, apart from tradition,
to assume that the theory can provide the sort of representational explanation we
are familiar with in a theory that is commutative or Boolean at the fundamental
level. The ‘sui generis’ quantum probabilities can’t be understood as quantifying
ignorance about the pre-measurement value of an observable, as in a Boolean
theory, but cash out in terms of ‘what you’ll find if you measure,’ which involves

2 See (Janssen 2009) for a defense of this view of special relativity contra Harvey Brown (2006).
202 J. Bub

considering the outcome, at the Boolean macrolevel, of manipulating a quantum

system in a certain way.
A quantum ‘measurement’ is not really the same sort of thing as a measurement
of a physical quantity of a classical system. It involves putting a microsystem, like
a photon, in a macroscopic environment, say a beamsplitter or an analyzing filter,
where the photon is forced to make an intrinsically random transition recorded as
one of a number of macroscopically distinct alternatives in a macroscopic device
like a photon detector. The registration of the measurement outcome at the Boolean
macrolevel is crucial, because it is only with respect to a suitable structure of
alternative possibilities that it makes sense to talk about an event as definitely
occurring or not occurring, and this structure—characterized by Boole as capturing
the ‘conditions of possible experience’ (Pitowsky 1994)—is a Boolean algebra.
In special relativity, Lorentz contraction is a kinematic effect of motion in a
non-Newtonian space-time. The contraction is consistent with a dynamical account,
but such an account takes the forces involved to be Lorentz covariant, which is
to say that the dynamics is assumed to have symmetries that respect Lorentz
contraction as a kinematic effect of relative motion. (By contrast, in Lorentz’s
theory, the contraction is explained as a dynamical effect in Newtonian space-
time.) Analogously, the loss of information in a quantum measurement—Bohr’s
‘irreducible and uncontrollable disturbance’—is a kinematic effect of any process
of gaining information of the relevant sort, irrespective of the dynamical processes
involved in the measurement process. A solution to the ‘small’ measurement
problem—a dynamical explanation for the effective emergence of classicality, i.e.,
Booleanity, at the macrolevel—would amount to a proof that the unitary quantum
dynamics is consistent with the kinematics, analogous to a proof that relativistic
dynamics is consistent with the kinematics.
We argued in ‘Two dogmas’ that the ‘small’ measurement problem can be
resolved as a consistency problem by considering the dynamics of the measurement
process and the role of decoherence in the emergence of an effectively classical
probability space of macro-events to which the Born probabilities refer. (The pro-
posal was that decoherence solves the ‘small’ measurement problem, not the ‘big’
measurement problem—decoherence does not provide a dynamical explanation of
how an indefinite outcome in a quantum superposition becomes definite in the
unitary evolution of a measurement process.) An alternative solution is suggested by
Pitowsky’s combinatorial treatment of macroscopic objects in quantum mechanics
(Pitowsky 2004). Entanglement witnesses are observables that distinguish between
separable and entangled states. An entanglement witness for an entangled state is an
observable whose expectation value lies in a bounded interval for a separable state,
but is outside this interval for the entangled state. Pitowsky showed, for a large
class of entanglement witnesses, that for composite systems where measurement
of an entanglement witness requires many manipulations of individual particles,
entangled states that can be distinguished from separable states become rarer and
rarer as the number of particles increases, and he conjectured that this is true in
general. If Pitowsky’s conjecture is correct, a macrosystem in quantum mechanics
can be characterized as a many-particle system for which the measure of the set of
8 ‘Two Dogmas’ Redux 203

entangled states that can be distinguished from separable states tends to zero. In this
sense, a macrosystem is effectively a commutative or Boolean system.
On the information-theoretic interpretation, quantum mechanics is a new sort of
non-representational theory for an indeterministic universe, in which the quantum
state is a bookkeeping device for keeping track of probabilities and probabilistic
correlations between intrinsically random events. Probabilities are defined with
respect to a single Boolean perspective, the Boolean algebra generated by the
‘pointer-readings’ of what Bohr referred to as the ‘ultimate measuring instruments,’
which are ‘kept outside the system subject to quantum mechanical treatment’ (Bohr
1939, pp. 23–24):
In the system to which the quantum mechanical formalism is applied, it is of course possible
to include any intermediate auxiliary agency employed in the measuring processes. . . . The
only significant point is that in each case some ultimate measuring instruments, like the
scales and clocks which determine the frame of space-time coordination—on which, in the
last resort, even the definition of momentum and energy quantities rest—must always be
described entirely on classical lines, and consequently be kept outside the system subject to
quantum mechanical treatment.

Bohr did not, of course, refer to Boolean algebras, but the concept is simply a
precise way of codifying a significant aspect of what Bohr meant by a description
‘on classical lines’ or ‘in classical terms’ in his constant insistence that (his
emphasis) (Bohr 1949, p. 209)
however far the phenomena transcend the scope of classical physical explanation, the
account of all evidence must be expressed in classical terms.

by which he meant ‘unambiguous language with suitable application of the termi-

nology of classical physics’—for the simple reason, as he put it, that we need to
be able ‘to tell others what we have done and what we have learned.’ Formally
speaking, the significance of ‘classical’ here as being able ‘to tell others what we
have done and what we have learned’ is that the events in question should fit together
as a Boolean algebra, so conforming to Boole’s ‘conditions of possible experience.’
The solution to the ‘small’ measurement problem as a consistency problem does
not show that unitarity is suppressed at a certain level of size or complexity, so
that the non-Boolean possibility structure becomes Boolean and quantum becomes
classical at the macrolevel. Rather, the claim is that the unitary dynamics of
quantum mechanics is consistent with the kinematics, in the sense that treating
the measurement process dynamically and ignoring certain information that is
in practice inaccessible makes it extremely difficult to detect the phenomena of
interference and entanglement associated with non-Booleanity. In this sense, an
effectively classical or Boolean probability space of Born probabilities can be
associated with our observations at the macrolevel.
Any system, of any complexity, is fundamentally a non-Boolean quantum system
and can be treated as such, in principle, which is to say that a unitary dynamical
analysis can be applied to whatever level of precision you like. But we see actual
events happen at the Boolean macrolevel. At the end of a chain of instruments
and recording devices, some particular system, M, functions as the ‘ultimate
204 J. Bub

measuring instrument’ with respect to which an event corresponding to a definite

measurement outcome occurs in an associated Boolean algebra, whose selection
is not the outcome of a dynamical evolution described by the theory. The system
M, or any part of M, can be treated quantum mechanically, but then some other
system, M  , treated as classical or commutative or Boolean, plays the role of
the ultimate measuring instrument in any application of the theory. The outcome
of a measurement is an intrinsically random event at the macrolevel, something
that actually happens, not described by the deterministic unitary dynamics, so
outside the theory. Putting it differently, the ‘collapse,’ as a conditionalization of the
quantum state, is something you put in by hand after recording the actual outcome.
The physics doesn’t give it to you.
As Pauli put it (Pauli 1954):
Observation thereby takes on the character of irrational, unique actuality with unpredictable
outcome. . . . Contrasted with this irrational aspect of concrete phenomena which are
determined in their actuality, there stands the rational aspect of an abstract ordering of
the possibilities of statements by means of the mathematical concept of probability and the
ψ-function.

A representational theory proposes a primitive ontology, perhaps of particles

or fields of a certain sort, and a dynamics that describes how things change over
time. The non-Boolean physics of quantum mechanics does not provide a repre-
sentational explanation of phenomena. Rather, a quantum mechanical explanation
involves showing how a Boolean output (a measurement outcome) is obtained
from a Boolean input (a state preparation) via a non-Boolean link. If a ‘world’
in which truth and falsity, ‘this’ rather than ‘that,’ makes sense is Boolean, then
there is no quantum ‘world,’ as Bohr is reported to have said,3 and it would
be misleading to attempt to picture it. In this sense, a quantum mechanical
explanation is ‘operational,’ but this is not simply a matter of convenience or
philosophical persuasion. We adopt quantum mechanics—the theoretical formalism
of the non-Boolean link between Boolean input and output—for empirical reasons,
the failure of classical physics to explain certain phenomena, and because there is
no satisfactory representational theory of such phenomena.
Quantum mechanics on the Everett interpretation is regarded, certainly by its
proponents, as a perfectly good representational theory: it explains phenomena in
terms of an underlying ontology and a dynamics that accounts for change over time.
As I will show in the following section, the Everett interpretation leads to modal
contradictions in scenarios that involve ‘encapsulated’ measurements.

3 Aage Petersen (Petersen 1963, p. 12): ‘When asked whether the algorithm of quantum mechanics
could be considered as somehow mirroring an underlying quantum world, Bohr would answer,
“There is no quantum world. There is only an abstract quantum physical description. It is wrong to
think that the task of physics is to find out how nature is. Physics concerns what we can say about
nature.”
8 ‘Two Dogmas’ Redux 205

8.3 Encapsulated Measurements and the Everett

Interpretation

According to the Born rule, the probability of finding the 

outcome a in a measure-
ment of an observable A on a system S in the state |ψ = a a|ψ|a is:

pψ (a) = Tr(Pa Pψ ) = |a|ψ|2 (8.1)

where Pa is the projection operator onto the eigenstate |a and Pψ is the projection
operator onto the quantum state. After the measurement, the state is updated to

|ψ −→ |a (8.2)

On the information-theoretic interpretation, the state update is understood as condi-

tionalization (with necessary loss of information) on the measurement outcome.
For a sequence of measurements, perhaps by two different observers, of A
followed by B on the same system S initially in the state |ψ, where A and B need
not commute, the conditional probability of the outcome b of B given the outcome
a of A is

pψ (a, b) pψ (a, b)
pψ (b|a) =  = = Tr(Pb Pa ) = |b|a|2 (8.3)
p
b ψ (a, b) pψ (a)

with pψ (a, b) = |a|ψ|2 · |b|a|2 = pψ (a) · |b|a|2 .

According to Everett’s relative state interpretation, a measurement is a unitary
transformation, or equivalently an isometry, V , that correlates the pointer-reading
state of the measuring instrument and the associated memory state of an observer
with the state of the observed system. (In the following, I’ll simply refer to the
memory state, with the understanding that this includes the measuring instrument
and recording device and any other systems involved in the measurement.) In the
V
case of a projective measurement by an observer O, the transformation HS −→
HS ⊗HO maps |aS onto |aS ⊗|Aa O , where |Aa O is the memory state correlated
with the eigenstate |aS , so

V 
|ψ −→ |Ψ  = a|ψ|aS ⊗ |Aa O (8.4)
a

Baumann and Wolf (2018) formulate the Born rule for the memory state |Aa O of
the observer O having observed a in the relative state interpretation, where there is
no commitment to the existence of a physical record of the measurement outcome as
classical information. The probability of the observer O observing a in this sense is

qψ (a) = Tr(1 ⊗ PAa · V Pψ V † ) (8.5)

206 J. Bub

where PAa is the projection operator onto the memory state |Aa O . This is
equivalent to the probability pψ (a) in the standard theory.
This equivalence can also be interpreted as showing the movability of the
notorious Heisenberg ‘cut’ between the observed system and the composite system
doing the observing, if the observing system is treated as a classical or Boolean
system. The cut can be placed between the system and the measuring instrument
plus observer, so that only the system is treated quantum mechanically, or between
the measuring instrument and observer, so that the system plus instrument is
treated as a composite quantum system. Similar shifts are possible if the measuring
instrument is subdivided into several component systems, e.g., if a recording device
is regarded as a separate component system, or if the observer is subdivided into
component systems involved in the registration of a measurement outcome.
On the relative state interpretation, the probability that S is in the state |a   but
the observer sees a is

qψ (|a  , a) = Tr(Pa  ⊗ PAa · V Pψ V † ) = δ a  ,a (8.6)

So the probability that S is in the state |a after O observes a is 1, as in the standard
theory according to the state update rule.
For a sequence of measurements of observables A and B by two observers, O1
and O2 , on the same system S, the conditional probability for the outcome b given
the outcome a is

qψ (a, b) Tr(1 ⊗ PAa ⊗ PBb · VO2 VO1 Pψ VO† 1 VO† 2 )

qψ (b|a) ==  = † †
b qψ (a, b) b Tr(1 ⊗ PAa ⊗ PBb · VO2 VO1 Pψ VO1 VO2 )
(8.7)

which turns out to be the same as the conditional probability given by the state
update rule: qψ (b|a) = pψ (b|a).
So Everett’s relative state interpretation and the information-theoretic interpre-
tation make the same predictions, both for the probability of an outcome of an
A-measurement on a system S, and for the conditional probability of an outcome of
a B-measurement, given an outcome of a prior A-measurement on the same system
S by two observers, O1 and O2 .
As Baumann and Wolf show, this is no longer the case for conditional probabil-
ities involving encapsulated measurements at different levels of observation, where
an observer O measures an observable of a system S and a super-observer measures
an arbitrary observable of the joint system S + O. In other words, the movability of
the cut is restricted to measurements that are not encapsulated. Note that both the
observer and the super-observer could be quantum automata, so the analysis is not
restricted to observers as conscious agents.
Suppose an observer O measures an observable with eigenstates |aS on a system
S in a state |ψS = a a|ψ|aS , and a super-observer SO then measures an
observable with eigenstates |bS+O in the Hilbert space of the composite system
S + O.
8 ‘Two Dogmas’ Redux 207

On the information-theoretic interpretation of the standard theory, the state of

S + O after the measurement by O is |a ⊗ Aa S+O = |aS ⊗ |Aa O , where |Aa O is
the memory state of O correlated with the outcome a. The probability of the super-
observer finding the outcome b given that the observer O obtained the outcome a is
then

pψ (b|a) = |b|a ⊗ Aa S+O |2 (8.8)

On the relative state interpretation, after the unitary evolutions associated with
the observer O’s measurement of the S-observable with eigenstates |a, and the
super-observer SO’s measurement of the (S + O)-observable with eigenstates |b,
the state of the composite system S + O + SO is

|Ψ  = a|ψb|a ⊗ Aa |bS+O |Bb SO (8.9)
a,b

where |Bb O is the memory state of SO correlated with the outcome b. The
probability of the super-observer SO finding the outcome b given that the observer
O obtained the outcome a is then qψ (b|a) as given by (8.7), with |Ψ Ψ | for
VO2 VO1 Pψ VO† 1 VO† 2 . The numerator of this expression, after taking the trace over
the (S + O)-space followed by the trace over the O-space of the projection onto
PBb followed by the projection onto PAa , becomes

pψ (b|a) a  |ψψ|a  b|a  ⊗ Aa  a  ⊗ Aa  |b (8.10)
a  ,a 

so

pψ (b|a) 
a ,a  a  |ψψ|a  b|a  ⊗ Aa  a  ⊗ Aa  |b
qψ (b|a) =  (8.11)
b Tr(1 ⊗ PAa ⊗ PBb · |Ψ Ψ |)

Baumann and Wolf point out that this is equal to pψ (b|a) in (8.8) only if |bS+O =
|aS ⊗ |Aa O for all b, i.e., if |bS+O is a product state of an S-state |aS and an
O-state |Aa O , which is not the case for general encapsulated measurements.
Here O’s measurement outcome states |Aa O are understood as relative to
states of SO, and there needn’t be a record of O’s measurement outcome as
classical or Boolean information. If there is a classical record of the outcome
then, as Baumann and Wolf show, SO’s conditional probability is the same as the
conditional probability in the standard theory.
To see that the relative state interpretation leads to a modal contradiction,
consider the Frauchiger-Renner modification of the ‘Wigner’s Friend’ thought
experiment (Frauchiger and Renner 2018). Frauchiger and Renner add a second
Friend (F ) and a second Wigner (W ). Friend F measures an observable with
eigenvalues ‘heads’ or ‘tails’ and eigenstates |h, |t on a system R in the state
208 J. Bub


√1 |h + 23 |t. One could say that F ‘tosses a biased quantum coin’ with
3
probabilities 1/3 for heads and 2/3 for tails. She prepares a qubit S in the state | ↓
if the outcome is h, or in the state | → = √1 (| ↓ + | ↑) if the outcome is t, and
2
sends S to F . When F receives S, he measures an S-observable with eigenstates
| ↓, | ↑ and corresponding eigenvalues − 12 , 12 . The Friends F and F are in two
laboratories, L and L, which are assumed to be completely isolated from each other,
except briefly when F sends the qubit S to F .
Now suppose that W is able to measure an L-observable with eigenstates

1
|failL = √ (|hL + |tL )
2
1
|okL = √ (|hL − |tL )
2

and W is able to measure an L-observable with eigenstates

1 1 1
|failL = √ (| − L + | L )
2 2 2
1 1 1
|okL = √ (| − L − | L )
2 2 2

Suppose W measures first and we stop the experiment at that point. On the
relative state interpretation, F ’s memory and all the systems in F ’s laboratory
involved in the measurement of R become entangled by the unitary transformation,
and similarly F ’s memory and all the systems in F ’s laboratory involved in the
measurement of S become entangled. If R is in the state |hR and F measures R,
the entire laboratory L evolves to the state |hL = |hR |hF , where |hF represents
the state of F ’s memory plus measuring and recording device plus all the systems
in F ’s lab connected to the measuring and recording device after the registration of
outcome h. Similarly for the state |tL , and for the states |− 12 L = |− 12 S |− 12 F and
| 12 L = | 12 S | 12 F of F ’s lab L with respect to the registration of the outcomes ± 12
of the spin measurement. So after the evolutions associated with the measurements
by F and F , the state of the two laboratories is

1 1 2 | − 12 L + | 12 L
|Ψ  = √ |hL | − L + |tL √ (8.12)
3 2 3 2

1 1 2
= √ |hL | − L + |t |failL (8.13)
3 2 3 L

According to the state (8.13), the probability is zero that W gets the outcome ‘ok’
for his measurement, given that F obtained the outcome t for her measurement.
8 ‘Two Dogmas’ Redux 209

Now suppose that W and W both measure, and W measures before W , as in

the Frauchiger-Renner scenario. What is the probability that W will get ‘ok’ at the
later time given that F obtained ‘tails’ at the earlier time? The global entangled
state at any time simply expresses a correlation between measurements—the time
order of the measurements is irrelevant. The two measurements by W and W could
be spacelike connected, and in that case the order in which the measurements are
carried out clearly can’t make a different to the probability.4
Since observers in differently moving reference frames agree about which events
occur, even if they disagree about the order of events, an event that has zero
probability in some reference frame cannot occur for any observer in any reference
frame.5 In a reference frame in which W measures before W the probability is zero
that W gets ‘ok’ if F gets ‘tails,’ because this must be the same as the probability if
we take W out of the picture. It follows that if F gets ‘tails,’ the ‘ok’ measurement
outcome event cannot occur in a reference frame in which W measures before W .
Think of it from F ’s perspective. If she is certain that she found ‘tails,’ she
can predict with certainty that W ’s later measurement (timelike connected to her
‘tails’ event) will not result in the outcome ‘ok.’ She wouldn’t (shouldn’t!) change
her prediction because of the possible occurrence of an event spacelike connected
to W ’s measurement—not if she wants to be consistent with special relativity. W
finding ‘ok’ is a zero probability event in the absolute future of F ’s prediction that
cannot occur for F given the ‘tails’ outcome of her R-measurement, and hence
cannot occur in any reference frame. The fact that W ’s measurement, an event
spacelike connected to W ’s measurement, might occur after her prediction doesn’t
support altering the prediction, even though it would make F ’s memory of the
earlier event (and any trace of the earlier event in F ’s laboratory L) indefinite, so
no observer could be in a position to check whether or not F obtained the earlier
outcome t when W gets ‘ok.’
The state (8.13) can be expressed as
√
1 1 1 3
|Ψ  = √ |okL |okL − √ |okL |failL + √ |failL |okL + |failL |failL
12 12 12 2
(8.14)
It follows that the probability that both Wigners obtain the outcome ‘ok’ is 1/12.
Since the two Wigners measure commuting observables on separate systems, W can
communicate the outcome ‘ok’ of her measurement to W , and her prediction that
she is certain, given the outcome ‘ok,’ that W will obtain ‘fail,’ without ‘collapsing’
the global entangled state. Then in a round in which W obtains the outcome ‘ok’ for
his measurement and so is certain that the outcome is ‘ok,’ he is also certain that the
outcome of his measurement is not ‘ok.’

4 Thanks to Renato Renner for pointing this out.

5 InBub and Stairs (2014), Allen Stairs and I proposed this as a consistency condition to avoid
potential contradictions in quantum interactions with closed timelike curves.
210 J. Bub

Frauchiger and Renner derive this modal contradiction on the basis of three
assumptions, Q, S, and C. Assumption Q (for ‘quantum’) says that an agent can
‘be certain’ of the value of an observable at time t if the quantum state assigns
probability 1 to this value after a measurement that is completed at time t, and
the agent has established that the system is in this state. An agent can also ‘be
certain’ that the outcome does not occur if the probability is zero. One should
also include under Q the assumption that measurements are quantum interactions
that transform the global quantum state unitarily. Assumption S (for ‘single value’)
says, with respect to ‘being certain,’ that an agent can’t ‘be certain’ that the value
of an observable is equal to v at time t and also ‘be certain’ that the value of this
observable is not equal to v at time t.6 Finally, assumption C (for ‘consistency’) is a
transitivity condition for ‘being certain’: if an agent ‘is certain’ that another agent,
reasoning with the same theory, ‘is certain’ of the value of an observable at time t,
then the first agent can also ‘be certain’ of this value at time t.
As used in the Frauchiger-Renner argument, ‘being certain’ is a technical term
entirely specified by the three assumptions. In other words, ‘being certain’ can mean
whatever you like, provided that the only inferences involving the term are those
sanctioned by the three assumptions. In particular, it is not part of the Frauchiger-
Renner argument that if an agent ‘is certain’ that the value of an observable is v at
time t, then the value of the observable is indeed v at time t—the observable might
have multiple values, as in the Everett interpretation. So it does not follow that the
proposition ‘the value of the observable is v at time t’ is true, or that the system has
the property corresponding to the value v of the observable at time t. Also, it is not
part of the Frauchiger-Renner argument that if an agent ‘is certain’ of the value of
an observable, then the agent knows the value, in any sense of ‘know’ that entails
the truth of what the agent knows.
The argument involves measurements, and inferences by the agents via the notion
of ‘being certain,’ that can all be modeled as unitary transformations to the agents’
memories. So picture the agents as quantum automata, evolving unitarily, where
these unitary evolutions result in changes to the global quantum state of the two
friends and the two Wigners that reflect changes in the agents’ memories. What
Frauchiger and Renner show is that these assumptions lead to memory changes that
end up in a modal contradiction.
Obviously, the relative state interpretation is inconsistent with the assumption
S in the sense that different branches of the quantum state can be associated
with different values of an observable after a measurement. So, with respect to a
particular branch, an agent can be certain that an observable has a certain value
and, with respect to a different branch, an agent associated with that branch can be
certain that the observable has a different value. This is not a problem. Rather, the

6 The standard axiom system for the modal logic of the operator ‘I am certain that’ includes the
axiom ‘if I am certain that A, then it is not the case that I am certain that not-A’ (assuming the
accessibility relation is serial and does not have any dead-ends). Thanks to Eric Pacuit for pointing
this out.
8 ‘Two Dogmas’ Redux 211

problem for the relative state interpretation is that there is a branch of the global
quantum state after W ’s measurement, which has a non-zero probability of 1/12,
with a memory entry that is inconsistent with S: W is certain that the outcome of his
measurement is ‘fail’ and also certain that the outcome is ‘ok.’
Renner gives a full description of the state of the global system as it evolves
unitarily with the measurements and inferential steps at every stage of the experi-
ment, which is repeated over many rounds until W and W both obtain the outcome
‘ok’ (Renner 2019). The measurements are assumed to be performed by the agents
F , F, W , W in sequence beginning at times 00, 10, 20, and 30 (and ending at times
01, 11, 21, and 31).
So, for example, in round n after F ’s measurement the global entangled state is
|Ψ n:11 (cf. (8.12)):

√1 |h |h, no conclusion | ↓S |z = − 12 F

3 R F

+ 13 |tR |t , so I am certain that w = fail at n:31F | ↓S |z = − 12 F

+ 13 |tR |t , so I am certain that w = fail at n:31F | ↑S |z = 12 F (8.15)

After the measurements by W and W with outcomes w̄ and w, the global entangled
state is |Ψ n:31 :

12 |okL |ok, so I am certain that w = failW |okL |I am certain that w = fail; I observe w = ok!W
1


− 12
1
|okL |ok, so I am certain that w = failW |failL |I am certain that w = fail; I observe w = failW

+ 12
1
|failL |fail, no conclusionW |okL |no conclusion previously; I observe w = okW
√
+ 2 |failL |fail, no conclusionW |failL |no conclusion previously; I observe w = failW
3
(8.16)

The inconsistency with S is apparent on the first branch.

8.4 A Clarification

Baumann and Wolf argue that standard quantum mechanics with the ‘collapse’ rule
for measurements, and quantum mechanics on Everett’s relative state interpretation,
are really two different theories, not different interpretations of the same theory. For
the Frauchiger-Renner scenario, they derive the conditional probability
212 J. Bub

pΨ (okW |tF ) = 0 (8.17)

for the standard theory by supposing that F updates the quantum state on the basis
of the outcome of her measurement of R and corresponding state preparation of S,
but F ’s measurement is described as a unitary transformation from F ’s perspective,
so the state of the two labs after the measurements by F (assuming F found ‘tails’)
and F is |tL (| − 12 L + | + 12 L ) = |tL |failL . For the relative state theory, they
derive the conditional probability

qΨ (okW |tF ) = 1/6 (8.18)

This would seem to be in conflict with the claim in the previous section that the
probability is zero that W gets the outcome ‘ok’ for his measurement, given that
F obtained the outcome ‘tails’ for her measurement, whether or not W measures
before W .
To understand the meaning of the probability qΨ (okW |tF ) as Baumann and Wolf
define it, note that after the unitary evolution associated with W ’s measurement on
the lab L, the state of the two laboratories and W is (from (8.14))

1 1
Φ = √ |okL |okW |okL − √ |okL |okW |failL
12 12
√
1 3
+ √ |failL |failW |okL + |failL |failW |failL (8.19)
12 2

which can be expressed as

1  5  3 1 
Φ = √ |hL √ |failW − √ |okW |failL
2 6 10 10
1  1 1 
+ √ √ |failW + √ |okW |okL
6 2 2
  
1 5 3 1
√ |tL √ |failW + √ |okW |failL
2 6 10 10
1  1 1 
+ √ √ |failW − √ |okW |okL (8.20)
6 2 2

Equation (8.20) is equivalent to (19) in (Healey 2018).

It follows from (8.20) that the joint probability qΦ (okW , tL ) is defined after W ’s
measurement and equals 1/12. Since qΦ (okW , tL )+qΦ (failW , tL ) = 1/12+5/12 =
1/2,

qΦ (okW , tL )
qΦ (okW |tL ) = = 1/6 (8.21)
qΦ (okW , tL ) + qΦ (failW , tL )
8 ‘Two Dogmas’ Redux 213

The conditional probability qΦ (okW |tL ) is derived from the jointly observable
statistics at the latest time in the unitary evolution, so at the time immediately after
W ’s measurement, following W ’s prior measurement.7 After W ’s measurement
of the observable with eigenvalues ‘ok,’ ‘fail,’, the L-observable with eigenvalues
‘heads,’ ‘tails’ is indefinite. The probability qΦ (okW |tL ) refers to a situation in
which a super-observer W measures an observable with eigenvalues ‘heads’ or
‘tails’ on L and notes the fraction of cases where W gets ‘ok’ to those in which this
measurement results in the outcome ‘tails.’ The ‘tails’ value in this measurement
is randomly related to the ‘tails’ outcome of F ’s previous measurement before W ’s
intervention, so the probability qΦ (okW |tL ), which Baumann and Wolf identify with
qΦ (okW |tF ), is not relevant to F ’s prediction that W will find ‘fail’ in the cases
where she finds ‘tails.’ Certainly, given the disruptive nature of W ’s measurement,
it is not the case that W will find ‘tails’ after W ’s measurement if and only if
F ’s measurement resulted in the outcome ‘tails’ at the earlier time before W ’s
measurement.8
This is not a critique of Baumann and Wolf, who make no such claim. The
purpose of this section is simply to clarify the difference between the conditional
probability qΦ (okW |tF ), as Baumann and Wolf define it, and F ’s conditional
prediction that W will find ‘fail’ in the cases where she finds ‘tails.’

8.5 Concluding Remarks

Frauchiger and Renner present their argument as demonstrating that any interpre-
tation of quantum mechanics that accepts assumptions Q, S, and C is inconsistent,
and they point out which assumptions are rejected by specific interpretations. For
the relative state interpretation and the many-worlds interpretation, they say that
these interpretations conflict with assumption S, because
any quantum measurement results in a branching into different ‘worlds,’ in each of which
one of the possible measurement outcomes occurs.

But this is not the issue. Rather, Frauchiger and Renner show something much more
devastating: that for a particular scenario with encapsulated measurements involving
multiple agents, there is a branch of the global quantum state with a contradictory
memory entry.
An interpretation of quantum mechanics is a proposal to reformulate quantum
mechanics as a Boolean theory in the representational sense, either by introducing
hidden variables, or by proposing that every possible outcome occurs in a mea-
surement, or in some other way. An implicit assumption of the Frauchiger-Renner
argument is that quantum mechanics is understood as a representational theory,

7 Thanks to Veronika Baumann for clarifying this.

8 Thanks to Renato Renner for this observation.
214 J. Bub

in the minimal sense that observers can be represented as physical systems, with
the possibility that observers can observe other observers. What the Frauchiger-
Renner argument really shows is that quantum mechanics can’t be interpreted as a
representational theory at all.

Acknowledgements Thanks to Veronika Baumann, Michael Dascal, Allen Stairs, and Tony
Sudbery for critical comments on earlier drafts.

References

Baumann, V., & Wolf, S. (2018). On formalisms and interpretations. Quantum, 2, 99–111.
Bell, J. (1987). Quantum mechanics for cosmologists. In J. S. Bell (Ed.), Speakable and
unspeakable in quantum mechanics. Cambridge: Cambridge University Press.
Bohr, N. (1939). The causality problem in atomic physics. In New theories in physics (pp. 11–30).
Warsaw: International Institute of Intellectual Cooperation.
Bohr, N. (1949). Discussions with Einstein on epistemological problems in modern physics. In
P. A. Schilpp (Ed.), Albert Einstein: Philosopher-scientist (The library of living philosophers,
vol. 7, pp. 201–241). Evanston: Open Court.
Brown, H. R. (2006). Physical relativity: Space-time structure from a dynamical perspective.
Oxford: Clarendon Press.
Bub, J. (2016). Bananaworld: Quantum mechanics for primates. Oxford: Oxford University Press.
Bub, J., & Pitowsky, I. (2010). Two dogmas about quantum mechanics. In S. Saunders, J. Barrett,
A. Kent, & D. Wallace (Eds.), Many worlds? Everett, quantum theory, and reality (pp. 431–
456). Oxford: Oxford University Press.
Bub, J., & Stairs, A. (2014). Quantum interactions with closed timelike curves and superluminal
signaling. Physical Review A, 89, 022311.
Frauchiger, D., & Renner, R. (2018). Quantum theory cannot consistently describe the use of itself.
Nature Communications, 9, article number 3711.
Gleason, A. N. (1957). Measures on the closed subspaces of Hilbert space. Journal of Mathematics
and Mechanics, 6, 885–893.
Healey, R. (2018). Quantum theory and the limits of objectivity. Foundations of Physics, 49, 1568–
1589.
Janssen, M. (2009). Drawing the line between kinematics and dynamics in special relativity.
Studies in History and Philosophy of Modern Physics, 40, 26–52.
Pauli, W. (1954). ‘Probability and physics,’ In C. P. Enz & K. von Meyenn (Eds.), Wolfgang Pauli:
Writings on physics and philosophy (p. 46). Berlin: Springer. The article was first published in
(1954) Dialectica, 8, 112–124.
Petersen, A. (1963). The philosophy of Niels Bohr. Bulletin of the Atomic Scientists, 19, 8–14.
Pitowsky, I. (1994). George Boole’s ‘conditions of possible experience’ and the quantum puzzle.
British Journal for the Philosophy of Science, 45, 95–125. ‘These . . . may be termed the
conditions of possible experience. When satisfied they indicate that the data may have, when
not satisfied they indicate that the data cannot have, resulted from actual observation.’ Quoted
by Pitowsky from George Boole, ‘On the theory of probabilities,’ Philosophical Transactions
of the Royal Society of London, 152, 225–252 (1862). The quotation is from p. 229.
Pitowsky, I. (2004). Macroscopic objects in quantum mechanics: A combinatorial approach.
Physical Review A, 70, 022103–1–6.
Pitowsky, I. (2007). Quantum mechanics as a theory of probability. In W. Demopoulos & I.
Pitowsky (Eds.), Festschrift in honor of Jeffrey Bub (Western Ontario Series in Philosophy
of Science). New York: Springer.
8 ‘Two Dogmas’ Redux 215

Renner, R. (2019). Notes on the discussion of ‘Quantum theory cannot consistently describe the
use of itself.’ Unpublished.
von Neumann, J. (1937). Quantum logics: Strict- and probability-logics. A 1937 unfinished
manuscript published in A. H. Taub (Ed.), Collected works of John von Neumann (vol. 4, pp.
195–197). Oxford/New York: Pergamon Press.
von Neumann, J. (1954). Unsolved problems in mathematics. An address to the international
mathematical congress, Amsterdam, 2 Sept 1954. In M. Rédei & M. Stöltzner (Eds.), John von
Neumann and the foundations of quantum physics (pp. 231–245). Dordrecht: Kluwer Academic
Publishers, 2001.
Chapter 9
Physical Computability Theses

B. Jack Copeland and Oron Shagrir

Abstract The Church-Turing thesis asserts that every effectively computable

function is Turing computable. On the other hand, the physical Church-Turing
Thesis (PCTT) concerns the computational power of physical systems, regardless
of whether these perform effective computations. We distinguish three variants of
PCTT – modest, bold and super-bold – and examine some objections to each. We
highlight Itamar Pitowsky’s contributions to the formulation of these three variants
of PCTT, and discuss his insightful remarks regarding their validity. The distinction
between the modest and bold variants was originally advanced by Piccinini (Br J
Philos Sci 62:733–769, 2011). The modest variant concerns the behavior of physical
computing systems, while the bold variant is about the behavior of physical systems
more generally. Both say that this behavior, when formulated in terms of some
mathematical function, is Turing computable. We distinguish these two variants
from a third – the super-bold variant – concerning decidability questions about the
behavior of physical systems. This says, roughly, that every physical aspect of the
behavior of physical systems – e.g., stability, periodicity – is decidable (i.e. Turing
computable). We then examine some potential challenges to these three variants,
drawn from relativity theory, quantum mechanics, and elsewhere. We conclude that
all three variants are best viewed as open empirical hypotheses.

Keywords Physical computation · Church-Turing thesis · Decidability in

physics · Computation over the reals · Relativistic computation · Computing the
halting function · Quantum undecidability

B. J. Copeland
Department of Philosophy, University of Canterbury, Christchurch, New Zealand
e-mail: jack.copeland@canterbury.ac.nz
O. Shagrir ()
Department of Philosophy, The Hebrew University of Jerusalem, Jerusalem, Israel

© Springer Nature Switzerland AG 2020 217

M. Hemmo, O. Shenker (eds.), Quantum, Probability, Logic,
Jerusalem Studies in Philosophy and History of Science,
https://doi.org/10.1007/978-3-030-34316-3_9
218 B. J. Copeland and O. Shagrir

9.1 Introduction

The physical Church-Turing Thesis (PCTT) limits the behavior of physical systems
to Turing computability. We will distinguish several versions of PCTT, and will
discuss some possible empirical considerations against these. We give special
emphasis to Itamar Pitowsky’s contributions to the formulation of the physical
Church-Turing Thesis, and to his remarks concerning its validity.
The important distinction between ‘modest’ and ‘bold’ variants of PCTT was
noted by Gualtiero Piccinini (2011). Modest variants concern only computing
systems, while bold variants concern the behavior of physical systems without
restriction. The literature contains numerous examples of both modest and bold
formulations of PCTT; e.g., bold formulations appear in Deutsch (1985) and
Wolfram (1985), and modest formulations in Gandy (1980) and Copeland (2000).
We will distinguish the modest and bold variants from a third variant of PCTT,
which we term “super-bold”. This variant goes beyond the other two in including
decidability questions within its scope, saying, roughly, that every physical aspect
of the behavior of any physical system – e.g., stability, periodicity – is Turing
computable.
Once the distinction between the modest, bold, and super-bold variants is drawn,
we will give three different formulations of PCTT: a modest version PCTT-M, a
bold version PCTT-B, and a super-bold version PCTT-S. We will then review some
potential challenges to these three versions, drawn from relativity theory, quantum
mechanics, and elsewhere. We will conclude that all three are to be viewed as open
empirical hypotheses.

9.2 Three Physicality Theses: Modest, Bold and Super-Bold

The issue of whether every aspect of the physical world is Turing computable was
raised by several authors in the 1960s and 1970s, and the topic rose to prominence
in the mid-1980s. In 1985, Stephan Wolfram formulated a thesis that he described
as “a physical form of the Church-Turing hypothesis”: this says that the universal
Turing machine can simulate any physical system (1985: 735, 738). Wolfram put it
as follows: “[U]niversal computers are as powerful in their computational capacities
as any physically realizable system can be, so that they can simulate any physical
system” (Wolfram 1985: 735). In the same year David Deutsch (who laid the
foundations of quantum computation) formulated a principle that he also called
“the physical version of the Church-Turing principle” (Deutsch 1985: 99). Other
formulations were advanced by Earman (1986), Pour-El and Richards (1989) and
others.
Pitowsky also formulated a version of CTTP, in his paper “The Physical Church
Thesis and Physical Computational Complexity” (Pitowsky 1990), based on his
9 Physical Computability Theses 219

1987 lecture in the Eighth Jerusalem Philosophical Encounter Workshop.1 He said:

“Wolfram has recently proposed a thesis—‘a physical form of the Church-Turing
thesis’—which maintains, among other things, that no non-recursive function is
physically computable” (1990: 86). The “other things” pertain to computational
complexity: Pitowsky interpreted Wolfram as also claiming that the universal Turing
machine efficiently simulates physical processes, and Pitowsky challenged this
further contention (see also Pitowsky 1996, 2002). We will not discuss issues
of computational complexity here (but see Copeland and Shagrir 2019 for some
relevant discussion of the so-called “Extended Church-Turing Thesis”).
Many have confused PCTT with the original Church-Turing Thesis, formulated
by Alonzo Church (1936) and Alan Turing (1936); see Copeland (2017) for
discussion of misunderstandings of the original thesis. It is now becoming better
understood that, by ‘computation’, both Church and Turing meant a certain human
activity, numerical computation; in their day, computation was done by rote-workers
called “computers”, or, more rarely, “computors” (see e.g. Turing 1947: 387, 391).
Pitowsky correctly emphasized that PCTT and the original form of the thesis are
very different:
It should be noted that Wolfram’s contention has nothing to do with the original Church’s
thesis. By ‘every computable function is recursive,’ Church meant that the best analysis of
our pre-analytic notion of ‘computation’ is provided by the precise notion of recursiveness.
Indeed one sometimes refers to Church’s thesis as ‘empirical,’ but the meaning of that
statement, too, has nothing to do with physics. (Pitowsky 1990: 86)

We will use the term physical to refer to systems whose operations are in accord
with the actual laws of nature. These include not only actually existing systems, but
also idealized physical systems (systems that operate in some idealized conditions),
and physically possible systems that do not actually exist, but that could exist, or
will exist, or did exist, e.g., in the universe’s first moments. (Of course, there is no
consensus about exactly what counts as an idealized or possible physical system,
but this is not our concern here.)
We start by formulating a modest version of PCTT:
Modest Physical Church-Turing Thesis (PCTT-M) Every function computed by any
physical computing system is Turing computable.

The functions referred to need not necessarily be defined over discrete values (e.g.,
integers). Many physical systems presumably operate on real-valued magnitudes;
and the same is true of physical computers, e.g., analog computers. The nervous
system too might have analog computing components. All this requires consid-
eration of computability over the reals. The extension of Turing computability to
real-valued domains (or non-denumerable domains more generally) was initiated
by Turing, who talked about real computable numbers in his (1936). Definitions of

1 Some papers from the Workshop were published in 1990, in a special volume of Iyyun. The
volume also contains papers by Avishai Margalit, Charles Parsons, Warren Goldfarb, William Tait,
and Mark Steiner.
220 B. J. Copeland and O. Shagrir

real-valued computable functions have been provided by Grzegorczyk (1955, 1957),

Lacombe (1955), Mazur (1963), Pour-El (1974), Pour-El and Richards (1989), and
others. The definitions are related to one another but are not equivalent. The central
idea behind the definitions is that a universal Turing machine can approximate (or
simulate) the values of a function over the reals, to any degree of accuracy. We
describe one of the definitions in Sect. 9.4.
Bold theses, on the other hand, omit the restriction to computing systems: they
concern all (finite) physical systems, whether computing systems or not. Piccinini
emphasized, correctly, that the bold versions proposed by different writers are
often “logically independent of one another”, and exhibit “lack of confluence”
(2011: 747–748). The following bold thesis is based on the theses put forward
independently by Wolfram and Deutsch (Wolfram 1985; Deutsch 1985):
Bold Physical Church-Turing Thesis (PCTT-B) Every finite physical system can be
simulated to any specified degree of accuracy by a universal Turing machine.

Pitowsky in fact interpreted Wolfram as advancing a modest version of the

thesis, namely “that no non-recursive function is physically computable” (1990:
86). However, this is because Pitowsky was treating every physical process as
computation; he said, for example: “According to this rather simple picture, the
planets in their orbits may be conceived as ‘performing computations’” (1990:
84). Under this assumption, according to which all physical processes are com-
puting processes, there is no difference at all between modest and bold versions.
Piccinini, on the other hand, sensibly distinguished between computational and non-
computational physical processes; he took it that the planets in their orbits do not
perform computations. Against the backdrop of this distinction, both Wolfram’s
formulation and Deutsch’s formulation are bold: they concern physical systems in
general and not just computing systems.
In a recent paper, we introduced a new, stronger, form of PCTT, the “super-
bold” form, here named PCTT-S (Copeland et al. 2018). (The entailments between
PCTT-S and PCTT-B and PCTT-M are: PCTT-S entails PCTT-B, and, since PCTT-
B entails PCTT-M, PCTT-S also entails PCTT-M.) Unlike bold versions, the
super-bold form concerns not only the ability of the universal Turing machine to
simulate the behavior of physical systems (to any required degree of precision),
but additionally concerns decidability questions about this behavior, questions that
go beyond the simulation (or prediction) of behavior. Pitowsky (1996) provided
some instructive examples of yes/no questions that reach beyond the simulation or
prediction of behavior:
There are, however, questions about the future that do not involve any specific time but refer
to all the future. For example: ‘Is the solar system stable?’, ‘Is the motion of a given system,
in a known initial state, periodic?’ These are typical questions asked by physicists and
involve (unbounded) quantification over time. Thus, the question of periodicity is: ‘Does
there exist some T such that for all times t, xi (T+ t) = xi (t) for i = 1, 2, . . . n? Similarly
the question concerning stability is: ‘Does there exist some D such that for all times t the
maximal distance between the particles does not exceed D?’. (1996: 163)
9 Physical Computability Theses 221

The physical processes involved in these scenarios – the motion and stability
of physical systems – may (so far as we know at present) be Turing computable,
in the sense that the motions of the planets may admit of simulation by a Turing
machine, to any required degree of accuracy. (Another way to put this is that
possibly a Turing machine could be used to predict the locations of the planets at
every specific moment.) Yet the answers to certain physical questions about physical
systems – e.g., whether (under ideal conditions) the system’s motion eventually
terminates – may nevertheless be uncomputable. The situation is similar in the case
of the universal Turing machine itself: the machine’s behavior (consisting of the
physical actions of the read/write head) is always Turing computable in the sense
under discussion, since the behavior is produced by the Turing machine’s program;
yet the answers to some yes/no questions about the behavior, such as whether or
not the machine halts given certain inputs, are not Turing computable. Undecidable
questions also arise concerning the dynamics of cellular automata and many other
idealized physical systems.
We express this form of the physical thesis as follows:
Super-Bold Physical Church-Turing Thesis (PCTT-S): Every physical aspect of the
behavior of any physical system is Turing computable (decidable).

Are these three physical versions of the Church-Turing Thesis true, or even well-
evidenced? We discuss the modest version first.

9.3 Challenging the Modest Thesis: Relativistic Computation

There have been several attempts to cook up idealized physical machines able
to compute functions that no Turing machine can compute. Perhaps the most
interesting of these are “supertask” machines—machines that complete infinitely
many computational steps in a finite span of time. Among such machines are
accelerating machines (Copeland 1998, Copeland and Shagrir 2011), shrinking
machines (Davies 2001), and relativistic machines (Pitowsky 1990; Hogarth 1994;
Andréka et al. 2009). Pitowsky proposed a relativistic machine in the 1987 lecture
mentioned earlier. At the same time, Istvan Németi also proposed a relativistic
machine, which we outline below.
The fundamental idea behind relativistic machines is intriguing: these machines
operate in spacetime structures with the property that the entire endless lifetime of
one participant is included in the finite chronological past of a second participant—
sometimes called “the observer”. Thus the first participant could carry out an endless
computation, such as calculating each digit of π, in what is a finite timespan from
the observer’s point of view, say 1 hour. Pitowsky described a setup with extreme
acceleration that nevertheless functions in accordance with Special Relativity. His
example is of a mathematician “dying to know whether Fermat’s conjecture is true
or false” (the conjecture was unproved back then). The mathematician takes a trip
in a satellite orbiting the earth, while his students (and then their students, and then
222 B. J. Copeland and O. Shagrir

their students . . . ) “examine Fermat’s conjecture one case after another, that is, they
take quadruples of natural numbers (x,y,z,n), with n≥3, and check on a conventional
computer whether xn + yn = zn ” (Pitowsky 1990: 83).
Pitowsky suggested that similar set-ups could be replicated by spacetime struc-
tures in General Relativity (now sometimes called Malament-Hogarth spacetimes).
Mark Hogarth (1994) pointed out the non-recursive computational powers of
devices operating in these spacetimes. More recently, Etesi and Németi (2002),
Hogarth (2004), Welch (2008), Button (2009), and Barrett and Aitken (2010) have
further explored the computational powers of such devices, within and beyond the
arithmetical hierarchy. In what follows, we describe a relativistic machine RM that
arguably computes the halting function (we follow Shagrir and Pitowsky (2003)).
RM consists of a pair of communicating Turing machines TA and TB : TA ,
the observer, is in motion relative to TB , a universal machine. When the input
(m,n)—asking whether the mth Turing machine (in some enumeration of the Turing
machines) halts or not, when started on input n—enters TA , TA first prints 0
(meaning “never halts”) in its designated output cell and then transmits (m,n) to TB .
TB simulates the computation performed by the mth Turing machine when started
on input n and sends a signal back to TA if and only if the simulation terminates. If
TA receives a signal from TB , it deletes the 0 it previously wrote in its output cell
and writes 1 there instead (meaning “halts”). After 1 hour, TA ’s output cell shows 1
if the mth Turing machine halts on input n and shows 0 if the mth machine does not
halt on n. Since RM is able to do this for any input pair (m,n), RM will deliver any
desired value of the halting function. There is further discussion of RM in Copeland
and Shagrir (2007).
Here we turn to the question of whether RM is a counterexample to PCTT-
M. This depends on whether RM is physical and on whether it really computes
the halting function. First, is RM physical? Németi and his colleagues provide the
most physically realistic construction, locating machines like RM in setups that
include huge slowly rotating Kerr-type black holes (Andréka et al. 2009). They
emphasize that the computation is physical in the sense that “the principles of
quantum mechanics are not violated” and RM is “not in conflict with presently
accepted scientific principles”; and they suggest that humans might “even build”
a relativistic computer “sometime in the future” (Andréka et al. 2009: 501).
Naturally, all this is controversial. John Earman and John Norton pointed out
that communication between the participants is not trivially achieved, due to
extreme blue-shift effects, including the possibility of the signal destroying the
receiving participant (Earman and Norton 1993). Subsequently, several potential
solutions to this signaling problem have been proposed; see Etesi and Németi
(2002), Németi and Dávid (2006) and Andréka et al. (2009: 508–9). An additional
potential objection is that infinitary computation requires infinite memory, and so
requires infinite computation space (Pitowsky 1990: 84). Another way of putting
the objection is that the infinitary computation requires an unbounded amount
of matter-energy, which seems to violate the basic principles of quantum gravity
(Aaronson 2005)—although Németi and Dávid (2006) offer a proposed solution to
this problem. We return to the infinite memory problem in a later section.
9 Physical Computability Theses 223

Second, does RM compute the halting function? The answer depends on what
is included under the heading “physical computation”. We cannot even summarize
here the diverse array of competing accounts of physical computation found in the
current literature. But we can say that RM computes in the senses of “compute”
staked out by several of these accounts: the semantic account (Shagrir 2006;
Sprevak 2010), the mechanistic account (Copeland 2000; Miłkowski 2013; Fresco
2014; Piccinini 2015), the causal account (Chalmers 2011), and the BCC (broad
conception of computation) account (Copeland 1997). According to all these
accounts, RM is a counterexample to the modest thesis if RM is physical; at the
very least, RM seems to show that non-Turing physical computation is logically
possible. However, if computation is construed as the execution of an algorithm in
the classical sense, then RM does not compute.

9.4 Challenging the Bold Thesis

PCTT-B says that the behavior of every physical system can be simulated (to any
required degree of precision) by a Turing machine. Speculation that there may be
physical processes whose behavior cannot be calculated by the universal Turing
machine stretches back over several decades (see Copeland 2002 for a survey).
The focus has been not so much on idealized constructions such as RM (which,
if physical, is a counterexample to PCTT-B, as well as PCTT-M, since PCTT-B
entails PCTT-M); rather, the focus has been on whether the mathematical equations
governing the dynamics of physical systems are or are not Turing computable.
Early papers by Scarpellini (1963), Komar (1964), and Kreisel (1965, 1967) raised
this question. Georg Kreisel stated “There is no evidence that even present day
quantum theory is a mechanistic, i.e. recursive theory in the sense that a recursively
described system has recursive behavior” (1967: 270). Roger Penrose (1989, 1994)
conjectured that some mathematical insights are non-recursive. Assuming that this
mathematical thinking is carried out by some physical processes in the brain, the
bold thesis must then be false. But Penrose’s conjecture is highly controversial.
Another challenge to the bold thesis derives from the postulate of genuine
physical randomness (as opposed to quasi-randomness). Church showed in 1940
that any infinite, genuinely random sequence is uncomputable (Church 1940: 134–
135). Some have argued that, under certain conditions relating to unboundedness,
PCTT-B is false in a universe containing a random element (to use Turing’s term
from his 1950: 445; see also Turing 1948: 416). A random element is a system that
generates random sequences of bits. It is argued that if physical systems include
systems capable of producing unboundedly many digits of an infinite random binary
sequence, then PCTT-B is false (Copeland 2000, 2004; Calude and Svozil 2008;
Calude et al. 2010; Piccinini 2011). One of us, Copeland, also argues that (again
under unboundedness conditions) a digital computer using a random element forms
a counterexample to PCTT-M (Copeland 2002). However, the latter claim, unlike
the corresponding claim concerning PCTT-B, depends crucially upon one’s account
224 B. J. Copeland and O. Shagrir

of computation—Shagrir denies that a digital computer with a random element

computes the generated uncomputable sequences. In any case, though, it is an open
question whether genuine random elements, able to generate unboundedly many
digits of random binary sequences, exist in the physical universe, or physically could
exist.
A further challenge to the bold thesis was formulated by Piccinini (2011). One of
his argument’s premises is: “if our physical theories are correct, most transforma-
tions of the relevant physical properties are transformations of Turing-uncomputable
quantities into one another” (2011: 748). Another premise is: “a transformation
of one Turing-uncomputable value into another Turing-uncomputable value is
certainly a Turing-uncomputable operation” (2011: 748–749). The observation
that in a continuous physical world, not all arithmetical operations are Turing
computable is certainly correct (Copeland 1997). Where x and y are uncomputable
real numbers, x + y is in general not Turing computable, since the inputs x and y
cannot be inscribed on a Turing machine’s tape (except in the special case where x
and y have been given proper names, e.g., where the halting number is named “τ”—
but since there are only countably many proper names, most Turing uncomputable
real numbers must remain nameless). If, therefore, the bold thesis is simply that
“Any physical process is Turing computable” (Piccinini 2011: 746), then the thesis
is indeed false in a continuous universe; as Piccinini argued, a Turing machine can
receive at most a denumerable number of different inputs, and so the falsity of the
bold thesis results from the cardinality gap between the physical functions, defined
over non-denumerable domains, and the Turing computable functions, defined over
denumerable domains.
However, this simple argument shows merely that Piccinini’s version of the bold
thesis is of little interest if the physical world is assumed to be continuous. Our
own version of the thesis is sensitive to these considerations and requires only that
physical processes be simulable to any specified degree of accuracy. Our version
of the thesis is responsive to an account of (Turing machine) computation over the
reals according to which—contra Piccinini’s second premiss—the transformation
of one Turing uncomputable value into another Turing uncomputable value can
be a (Turing machine) computable operation. Relative to this account, the real-
valued functions of plus, identity, and inverse are computable by Turing machine,
even though these functions sometimes map Turing uncomputable inputs to Turing
uncomputable outputs: the definitions of real computable functions impose a
continuity constraint, enabling the approximation (simulation) of uncomputable
arguments and values. There are several (non-equivalent) characterizations of this
continuity constraint; for the purpose of illustration, we select the characterization
given by Andrzej Grzegorczyk (1955, 1957), and we adapt the exposition of Earman
(1986).
We start with numbers:

Definition 1: A sequence of rational numbers {xn } is said to be effectively computable

if there exist three Turing computable functions (over N) a,b,c such that
xn = (–1)c(n) a(n)÷b(n).
9 Physical Computability Theses 225

Definition 2: A real number r is said to be effectively computable if there is an effectively

computable sequence of rational numbers that converges effectively to r. (‘Converges
effectively’ means that there is an effectively computable function d over N such that
|r – xn | < 1÷2m whenever n ≥ d(m).)

Now to functions:

Definition 3: A function f is an effectively computable function over the reals if:

(i) f is sequentially computable, i.e., for each effectively computable sequence {rn } of
reals {f (rn )} is also effectively computable;
(ii) f is effectively uniformly continuous on rational intervals, i.e., if {xn } is an effective
enumeration of the rationals without repetitions then there is a three-place Turing
computable function g such that |f (r) – f (r )| < 1÷2k whenever xm < r, r < xn and
|r – r | < 1÷g(m,n,k) for all r,r ∈ R and all m,n,k ∈ N.
(If we confine f to a closed and bounded interval with computable end points then the above
definition simplifies: no enumeration is necessary and g is only a function of k.)

Importantly, given that plus is effectively uniformly continuous on rational intervals,

plus is a computable function, even though plus maps Turing uncomputable reals
to Turing uncomputable reals.
It is interesting that, by and large, known physical laws give rise to functions
over the reals that are computable in the sense just defined. A well-known exception
was discovered by Marian Pour-El and Ian Richards (1981), who showed that the
wave equation produces non-computable sequences for some computable initial
conditions (i.e., for some computable input sequences). In that respect, the wave
equation violates condition (i) in the definition of an effectively computable
function over the reals (Definition 3). Clearly, this exception forms a potential
counterexample to (our version of) the bold thesis.
However, this result of Pour-El and Richards is at the purely mathematical level,
not the physical level. In his discussion of Pour-El and Richards, Pitowsky (1990)
argued that their result does not refute the bold thesis, for two reasons:
Firstly, the function f in the initial condition, though a recursive real function, is an
extremely complex function. One can hardly expect such an initial condition to arise
‘naturally’ in any real physical situation. Secondly, we deal with recursive real functions,
and in physics we never get beyond a few decimal digits of accuracy anyway. (Pitowsky
1990: 86–87)

Nevertheless, the result does demonstrate that being recursive

is not a natural physical property. Physical processes do not necessarily preserve it.
(Pitowsky 1990: 87)

Even if, in our world, the initial conditions envisaged in the Pour-El & Richards
example do not occur, nevertheless these conditions could occur in some other
physically possible world in which the wave equation holds, so showing, as
Pitowsky said, that physical processes do not necessarily preserve recursiveness.
226 B. J. Copeland and O. Shagrir

9.5 Challenging the Super-Bold Thesis

We will turn next to PCTT-S. It might be objected that PCTT-S is immediately

false, as may be shown by considering a universal Turing machine implemented as
a physical system: many interesting questions about such a system are undecidable.
Pitowsky (1996) describes one such construction, due to Moore (1990), in which a
universal machine is realized in a moving particle bouncing between parabolic and
linear mirrors in a unit square. This system can certainly be simulated by a Turing
machine (since it is a Turing machine). But there are nevertheless undecidable
questions about its behavior:
At this stage we can apply Turing’s theorem on the undecidability of the halting problem. It
says: There is no algorithm to decide whether a universal Turing machine halts on a given
input. Translating this assertion into physical language, it means that there is no algorithm to
predict whether the particle ever enters the subset of the square corresponding to the halting
state. This assertion is valid when we know the exact initial conditions with unbounded
accuracy, or even with actually infinite accuracy. Therefore, to answer the question: ‘Is the
particle ever going to reach this region of space?’, Laplace’s Demon needs computational
powers exceeding any algorithm. In other words, he needs to consult an oracle. (Pitowsky
1996: 171)

Pitowsky further noted that many other yes/no questions about the system are
computationally undecidable. In fact, it follows from Rice’s theorem that “almost
every conceivable question about the unbounded future of a universal Turing
machine turns out to be computationally undecidable” (Pitowsky 1996: 171; see
also Harel and Feldman 1992). Rice’s theorem says: any nontrivial property about
the language recognized by a Turing machine is undecidable, where a property is
nontrivial if at least one Turing machine has the property and at least one does not.
However, the objection that PCTT-S is straight-out false can hardly be sustained.
It would by no means be facile to implement a Turing machine, with its infinite
tape, in a finite physical object. The assumption that the physical universe is able
to supply an infinite memory tape is controversial. If the number of particles in
the cosmos is finite, and each cell of the tape requires at least one particle for its
implementation, then clearly the assumption of an infinite physical tape must be
false. Against this, it might be pointed out that there are a number of well-known
constructions for shoehorning an infinite amount of memory into a finite object. For
example, a single strip of tape 1 metre long can be used: the first cell occupies the
first half-metre, the second cell the next quarter-metre, the third cell the next eighth
of a metre, and so on. However, this observation does not help matters. Obviously,
this construction can be realized only in a universe whose physics allows for the
infinite divisibility of the tape—again by no means an evidently true assumption.
Another way to implement the Turing machine with its infinite tape is in the
continuous (or rational) values of a physical magnitude, as described in Moore’s
system. Assume that each (potential) configuration of the Turing machine is
encoded in a (potentially infinite) sequence of 0s and 1s. The goal now is to
efficiently realize each sequence in a unique location of the particle in the (finite)
unit square. Much like the example of the 1 metre strip of tape above, this realization
9 Physical Computability Theses 227

can be achieved if we use potentially infinitely many different locations (x,y) within
a unit square, where x and y are continuous (or rational) values between 0 and 1.
This realization, however, requires that the (idealized) mirrors have the ability to
bounce the particle, accurately, into potentially infinitely many different positions
within the unit square. Moreover, if Laplace’s Demon wants to predict the future
position of the particle after k rounds, the Demon will have to be able to measure the
differences between arbitrarily close positions. As Pitowsky notes: “[T]o predict the
particle position with accuracy of 1/2 the demon has to measure initial conditions
with accuracy 2−(k+1) . The ratio of final error to initial error is 2k , and growing
exponentially with time k (measured by the number of rounds)” (1996, 167). This
means that implementing a Turing machine in Moore’s system requires determining
a particle’s position with practically infinite precision (an accuracy of 2−(k + 1) for
unbounded k), and it is questionable whether this implementation is physically
feasible. At any rate, the claim that this is physically feasible is, again, far from
obvious.
To summarize this discussion, PCTT-S hypothesizes in part that the universe
is physically unable to supply an infinite amount of memory, since if PCTT-S
is true, the resources for constructing a universal computing machine must be
unavailable (the other necessary resources, aside from the infinite memory, being
physically undemanding). This point helps illuminate the relationships between
PCTT-S and PCTT-M. Returning to the discussion of the relativistic machine RM
and the infinite memory problem raised in Sect. 9.3, it is clearly the case that, since
RM requires infinite memory, PCTT-S rules out RM (and this is to be expected,
since PCTT-M rules out RM, and PCTT-S entails PCTT-M). Nevertheless, the
falsity of PCTT-S, and the availability of infinite memory, would be insufficient
to falsify PCTT-M—a universe that consists of nothing but a universal Turing
machine with its infinite tape does not falsify PCTT-M. Thus, counterexamples
to PCTT-M must postulate not only the availability of infinite memory but also
additional physical principles of some sort, such as gravitational time dilation
or unbounded acceleration or unbounded shrinking of components. Relativistic,
accelerating and shrinking machines arguably invoke these principles successfully,
and, hence, provide counterexamples to PCTT-M.
Moving on to challenges to PCTT-S at the quantum level, there are undecidable
questions concerning the behavior of quantum systems. In 1986, Robert Geroch
and James Hartle argued that undecidable physical theories “should be no more
unsettling to physics than has the existence of well-posed problems unsolvable by
any algorithm have been to mathematics”; and they suggested such theories may
be “forced upon us” in the quantum domain (Geroch and Hartle 1986: 534, 549).
Arthur Komar raised “the issue of the macroscopic distinguishability of quantum
states” in 1964, claiming there is no effective procedure “for determining whether
two arbitrarily given physical states can be superposed to show interference effects”
(Komar 1964: 543–544). More recently, Jens Eisert, Markus Müller and Christian
Gogolin showed that “the very natural physical problem of determining whether
certain outcome sequences cannot occur in repeated quantum measurements is
undecidable, even though the same problem for classical measurements is readily
228 B. J. Copeland and O. Shagrir

decidable” (Eisert et al. 2012: 260501-1). (This is an example of a problem that

refers unboundedly to the future, but not to any specific time, as in Pitowsky’s
examples mentioned earlier.) Eisert, Müller and Gogolin went on to suggest that
“a plethora of problems” in quantum many-body physics and quantum computing
may be undecidable (2012: 260501-1 – 260501-4).
Dramatically, a 2015 Nature article by Toby Cubitt, David Perez-Garcia, and
Michael Wolf outlined a proof that “the spectral gap problem is algorithmically
undecidable: there cannot exist any algorithm that, given a description of the local
interactions, determines whether the resultant model is gapped or gapless” (Cubitt
et al. 2015: 207). Cubitt describes this as the “first undecidability result for a major
physics problem that people would really try to solve” (in Castelvecchi 2015).
The spectral gap, an important determinant of a material’s properties, refers to
the energy spectrum immediately above the ground energy level of a quantum many-
body system (assuming that a well-defined least energy level of the system exists);
the system is said to be gapless if this spectrum is continuous and gapped if there
is a well-defined next least energy level. The spectral gap problem for a quantum
many-body system is the problem of determining whether the system is gapped or
gapless, given the finite matrices describing the local interactions of the system.
In their proof, Cubitt et al. encode the halting problem in the spectral gap
problem, so showing that the latter is at least as hard as the former. The proof
involves an infinite family of 2-dimensional lattices of atoms; but they point out
that their result also applies to finite systems whose size increases: “Not only can
the lattice size at which the system switches from gapless to gapped be arbitrarily
large, the threshold at which this transition occurs is uncomputable” (Cubitt et al.
2015: 210–211). Their proof offers an interesting countermodel to the super-bold
thesis. The countermodel involves a physically relevant example of a finite system,
of increasing size, that lacks a Turing computable procedure for extrapolating future
behavior from (complete descriptions of) its current and past states.
It is debatable whether any of these quantum models matches the real quantum
world. Cubitt et al. admit that the model used in their proof is highly artificial,
saying “Whether the results can be extended to more natural models is yet to be
determined” (Cubitt et al. 2015: 211). There is also the question of whether the
spectral gap problem could become computable when only local Hilbert spaces
of realistically low dimensionality are considered. Nevertheless, these results are
certainly suggestive. The super-bold thesis cannot be taken for granted—even in a
finite quantum universe.

9.6 Conclusion

We have distinguished three theses about physical computability, and have discussed
some empirical evidence that might challenge these. Concerning the boldest
versions, PCTT-B and PCTT-S, both are false if the physical universe permits
infinite memory and genuine randomness. Even assuming that the physical universe
9 Physical Computability Theses 229

is deterministic, the most that can be said for PCTT-B and PCTT-S is that, to date,
there seems to be no decisive empirical evidence against them. PCTT-B and PCTT-
S are both thoroughly empirical theses; but matters are more complex in the case
of the modest thesis PCTT-M, since a conceptual issue also bears on the truth or
falsity of this thesis (even in a universe containing genuine randomness)—namely,
the difficult issue of what counts as physical computation. Our conclusion is that,
at the present stage of physical enquiry, it is unknown whether any of the theses is
true.

Acknowledgements An early version of this chapter was presented at the Workshop on Physics
and Computation (UCNC2018) in Fontainebleau, at the International Association for Computing
and Philosophy meeting (IACAP2018) in Warsaw, and at the research logic seminar in Tel Aviv
University. Thanks to the audiences for helpful discussion. We are also grateful to Piccinini for
his comments on a draft of this paper. Shagrir’s research was supported by the Israel Science
Foundation grant 830/18.

References

Aaronson, S. (2005). Guest column: NP-complete problems and physical reality. ACM SIGACT
News, 36(1), 30–52.
Andréka, H., Németi, I., & Németi, P. (2009). General relativistic hypercomputing and foundation
of mathematics. Natural Computing, 8, 499–516.
Barrett, J. A., & Aitken, W. (2010). A note on the physical possibility of ordinal computation. The
British Journal for the Philosophy of Science, 61, 867–874.
Button, T. (2009). SAD computers and two versions of the Church-Turing thesis. The British
Journal for the Philosophy of Science, 60, 765–792.
Calude, C. S., Dinneen, M. J., Dumitrescu, M., & Svozil, K. (2010). Experimental evidence of
quantum randomness incomputability. Physical Review A, 82: 022102–1-022102–8.
Calude, C. S., & Svozil, K. (2008). Quantum randomness and value indefiniteness. Advanced
Science Letters, 1, 165–168.
Castelvecchi, D. (2015). Paradox at the heart of mathematics makes physics problem unanswer-
able. Nature, 528, 207.
Chalmers, D. J. (2011). A computational foundation for the study of cognition. Journal of
Cognitive Science, 12, 323–357.
Church, A. (1936). An unsolvable problem of elementary number theory. American Journal of
Mathematics, 58, 345–363.
Church, A. (1940). On the concept of a random sequence. American Mathematical Society Bulletin,
46, 130–135.
Copeland, B. J. (1997). The broad conception of computation. American Behavioral Scientist, 40,
690–716.
Copeland, B. J. (1998). Even Turing machines can compute uncomputable functions. In C. Calude,
J. Casti, & M. Dinneen (Eds.), Unconventional models of computation. New York: Springer.
Copeland, B. J. (2000). Narrow versus wide mechanism. Journal of Philosophy 96, 5–32. Reprinted
in M. Scheutz (Ed.) (2002). Computationalism. Cambridge, MA: MIT Press.
Copeland, B. J. (2002). Hypercomputation. Minds and Machines, 12, 461–502.
Copeland, B. J. (2004). Hypercomputation: Philosophical issues. Theoretical Computer Science,
317, 251–267.
230 B. J. Copeland and O. Shagrir

Copeland, B. J. (2017). The Church-Turing thesis. In E. N. Zalta (Ed.), The Stanford encyclopedia
of philosophy. https://plato.stanford.edu/archives/win2017/entries/church-turing/
Copeland, B. J., & Shagrir, O. (2007). Physical computation: How general are Gandy’s principles
for mechanisms? Minds and Machines, 17, 217–231.
Copeland, B. J., & Shagrir, O. (2011). Do accelerating Turing machines compute the uncom-
putable? Minds and Machines, 21, 221–239.
Copeland, B. J., & Shagrir, O. (2019). The Church–Turing thesis—Logical limit, or breachable
barrier? Communications of the ACM, 62, 66–74.
Copeland, B. J., Shagrir, O., & Sprevak, M. (2018). Zuse’s thesis, Gandy’s thesis, and Penrose’s
thesis. In M. Cuffaro & S. Fletcher (Eds.), Physical perspectives on computation, computa-
tional perspectives on physics (pp. 39–59). Cambridge: Cambridge University Press.
Cubitt, T. S., Perez-Garcia, D., & Wolf, M. M. (2015). Undecidability of the spectral gap. Nature,
528(7581), 207–211.
Davies, B. E. (2001). Building infinite machines. The British Journal for the Philosophy of Science,
52, 671–682.
Deutsch, D. (1985). Quantum theory, the Church-Turing principle and the universal quantum
computer. Proceedings of the Royal Society of London A: Mathematical, Physical and
Engineering Sciences, 400(1818), 97–117.
Earman, J. (1986). A primer on determinism. Dordrecht: Reidel.
Earman, J., & Norton, J. D. (1993). Forever is a day: Supertasks in Pitowsky and Malament-
Hogarth spacetimes. Philosophy of Science, 60, 22–42.
Eisert, J., Müller, M. P., & Gogolin, C. (2012). Quantum measurement occurrence is undecidable.
Physical Review Letters, 108(26), 260501.
Etesi, G., & Németi, I. (2002). Non-Turing computations via Malament-Hogarth space-times.
International Journal of Theoretical Physics, 41, 341–370.
Fresco, N. (2014). Physical computation and cognitive science. Berlin: Springer.
Gandy, R. O. (1980). Church’s thesis and principles of mechanisms. In S. C. Kleene, J. Barwise,
H. J. Keisler, & K. Kunen (Eds.), The Kleene symposium. Amsterdam: North-Holland.
Geroch, R., & Hartle, J. B. (1986). Computability and physical theories. Foundations of Physics,
16, 533–550.
Grzegorczyk, A. (1955). Computable functionals. Fundamenta Mathematicae, 42, 168–203.
Grzegorczyk, A. (1957). On the definitions of computable real continuous functions. Fundamenta
Mathematicae, 44, 61–71.
Harel, D., & Feldman, Y. A. (1992). Algorithmics: The spirit of computing (3rd ed.). Harlow:
Addison-Wesley.
Hogarth, M. (1994). Non-Turing computers and non-Turing computability. Proceedings of the
Biennial Meeting of the Philosophy of Science Association, 1, 126–138.
Hogarth, M. (2004). Deciding arithmetic using SAD computers. The British Journal for the
Philosophy of Science, 55, 681–691.
Komar, A. (1964). Undecidability of macroscopically distinguishable states in quantum field
theory. Physical Review, second series, 133B, 542–544.
Kreisel, G. (1965). Mathematical logic. In T. L. Saaty (Ed.), Lectures on modern mathematics (Vol.
3). New York: Wiley.
Kreisel, G. (1967). Mathematical logic: What has it done for the philosophy of mathematics? In R.
Schoenman (Ed.), Bertrand Russell: Philosopher of the century. London: Allen and Unwin.
Lacombe, D. (1955). Extension de la notion de fonction récursive aux fonctions d’une ou plusieurs
variables réelles III. Comptes Rendus Académie des Sciences Paris, 241, 151–153.
Mazur, S. (1963). Computational analysis. Warsaw: Razprawy Matematyczne.
Miłkowski, M. (2013). Explaining the computational mind. Cambridge: MIT Press.
Moore, C. (1990). Unpredictability and undecidability in dynamical systems. Physical Review
Letters, 64, 2354–2357.
Németi, I., & Dávid, G. (2006). Relativistic computers and the Turing barrier. Journal of Applied
Mathematics and Computation, 178, 118–142.
9 Physical Computability Theses 231

Penrose, R. (1989). The emperor’s new mind: Concerning computers, minds and the laws of
physics. Oxford: Oxford University Press.
Penrose, R. (1994). Shadows of the mind. Oxford: Oxford University Press.
Piccinini, G. (2011). The physical Church-Turing thesis: Modest or bold? The British Journal for
the Philosophy of Science, 62, 733–769.
Piccinini, G. (2015). Physical computation: A mechanistic account. Oxford: Oxford University
Press.
Pitowsky, I. (1990). The physical Church thesis and physical computational complexity. Iyyun, 39,
81–99.
Pitowsky, I. (1996). Laplace’s demon consults an oracle: The computational complexity of
prediction. Studies in History and Philosophy of Science Part B, 27(2), 161–180.
Pitowsky, I. (2002). Quantum speed-up of computations. Philosophy of Science, 69, S168–S177.
Pour-El, M. B. (1974). Abstract computability and its relation to the general purpose analog
computer (some connections between logic, differential equations and analog computers).
Transactions of the American Mathematical Society, 199, 1–28.
Pour-El, M. B., & Richards, I. J. (1981). The wave equation with computable initial data such that
its unique solution is not computable. Advances in Mathematics, 39, 215–239.
Pour-El, M. B., & Richards, I. J. (1989). Computability in analysis and physics. Berlin: Springer.
Scarpellini, B. (1963). Zwei unentscheidbare Probleme der Analysis. Zeitschrift für mathematische
Logik und Grundlagen der Mathematik, 9, 265–289. English translation in Minds and Machines
12 (2002): 461–502.
Shagrir, O. (2006). Why we view the brain as a computer. Synthese, 153, 393–416.
Shagrir, O., & Pitowsky, I. (2003). Physical hypercomputation and the Church–Turing thesis.
Minds and Machines, 13, 87–101.
Sprevak, M. (2010). Computation, individuation, and the received view on representation. Studies
in History and Philosophy of Science, 41, 260–270.
Turing, A. M. (1936). On computable numbers, with an application to the Entscheidungsproblem.
Proceedings of the London Mathematical Society, Series, 2(42), 230–265.
Turing, A. M. (1947). Lecture on the Automatic Computing Engine. In Copeland, B. J. (2004). The
essential Turing: Seminal writings in computing, logic, philosophy, artificial intelligence, and
artificial life, plus the secrets of Enigma. Oxford University Press.
Turing, A. M. (1948). Intelligent machinery. In The essential Turing. Oxford University Press.
Turing, A.M. (1950). Computing machinery and intelligence. In The essential Turing. Oxford
University Press.
Welch, P. D. (2008). The extent of computation in Malament-Hogarth spacetimes. The British
Journal for the Philosophy of Science, 59, 659–674.
Wolfram, S. (1985). Undecidability and intractability in theoretical physics. Physical Review
Letters, 54, 735–738.
Chapter 10
Agents in Healey’s Pragmatist Quantum
Theory: A Comparison with Pitowsky’s
Approach to Quantum Mechanics

Mauro Dorato

Abstract In a series of related papers and in a book (Healey R. The quantum

revolution in philosophy. Oxford University Press, Oxford, 2017b), Richard Healey
has proposed and articulated a new pragmatist approach to quantum theory that I
will refer to as PQT. After briefly reviewing PQT by putting it into a more general
philosophical context, I will discuss Healey’s self-proclaimed quantum realism by
stressing that his agent-centered approach to quantum mechanics makes his position
rather close to a sophisticated form of instrumentalist, despite his claims to the
contrary. My discussion of the possible sources of incompatibility between PQT
and the view that agents can be regarded as physical systems (agent physicalism)
will allow me to compare Healey’s view of quantum theory with Pitowsky’s. In
particular, by focusing on the measurement problem, I will discuss the role of
observers and the notion of information in their respective philosophical approach.

Keywords Agents · Healey’s pragmatist quantum theory · Physicalism ·

Pitowski

10.1 Introduction

In a series of related papers (2012a, b, 2015, Healey 2017c, 2018a, 2018b) and
in a book (2017b), Richard Healey has proposed and articulated a new pragmatist
approach to quantum theory (PQT henceforth) that so far has not received the
attention it deserves. And yet, Healey’s approach – besides trying to clarify and
solve (or dissolve) well-known problems in the foundations of quantum mechanics
from a technical point of view – has relevant consequences also for philosophy in
general, in particular in areas like scientific realism, metaphysics, the philosophy of

M. Dorato ()
Department of Philosophy Communication and Performing Arts, Università degli Studi Roma 3,
Rome, Italy
e-mail: mauro.dorato@uniroma3.it

© Springer Nature Switzerland AG 2020 233

M. Hemmo, O. Shenker (eds.), Quantum, Probability, Logic,
Jerusalem Studies in Philosophy and History of Science,
https://doi.org/10.1007/978-3-030-34316-3_10
234 M. Dorato

probability, and the philosophy of language. Given that quantum theory (even in its
non-relativistic form) is one of two pillars of contemporary physics,1 any attempt
to offer a coherent and general outlook on the physical world that, like his, is not
superficial or merely suggestive, is certainly welcome.
The main aim of this paper is to show that, as a consequence of this ambition,
PQT raises important questions related to its compatibility with a narrowly defined
form of physicalism that enables us to clearly specify in what sense agents can
be regarded as quantum systems. It is important to note from the start that this
problem is related to the question whether quantum theory can be applied to all
physical systems – and therefore also to agents who use it (see Frauchiger and
Renner 2018; Healey 2018b) – to the extent that this applicability is a necessary
condition to consider them as physical systems. In what follows, I will not discuss
no-go arguments of this kind, since in the relationist context that Healey at least
initially advocated and that will be the subject of my discussion, he himself grants
that agents can be regarded as quantum systems by other agents.2
In this perspective, let me refer to this particular form of physicalism as “agent
physicalism”, in order to distinguish it from other approaches to physicalism that
Healey does not explicitly reject – among these, supervenience physicalism regarded
as the aim of physics3 - and that therefore will not be the target of my paper. In fact,
Healey grants that a particular form of physicalism – according to which our most
fundamental theory provides a complete ontology on which everything else must
supervene – is incompatible with PQT. From the outset, he claims that quantum
theory “lacking beables of its own, has no physical ontology and states no fact
about physical objects and events”, i.e. is not a fundamental description of the
physical world, but it rather amounts to a series of effective prescriptions yielding
correct attributions of quantum states to physical systems (Healey 2017b, pp. 11–
12). PQT’s normative approach to quantum theory is evident from the terminology
that in his papers abound (“prescribes”, “provides good advice”, etc.) and that
presupposes rules or norms that rational agents should follow). PQT postulates
in fact non-representational relations between physically situated agents’ epistemic
states (inter-subjectively shared credences) on the one hand, and mind-independent
probabilistic correlations instantiated by physical systems on the other. As we will

1 The other being, of course, general relativity. In this paper, I will neglect any considerations
emerging from quantum field theory to which, in the intention of his proposer, PQT certainly
applies. This fact is also clear from the examples that he discusses.
2 There are various forms of physicalism in the literature and here I am not interested in discussing

them all.
3 “Quine (1981, p. 98) offered this argument for a nonreductive physicalism: “Nothing happens

in the world, not the flutter of an eyelid, not the flicker of a thought, without some redistribution
of microphysical states. . . If the physicist suspected there was any event that did not consist in a
redistribution of the elementary states allowed for by his physical theory, he would seek a way of
supplementing his theory” (Healey 2017a, 235). Even though Healey never rejected supervenience
physicalism, he adds that: “If anything, the quantum revolution set back progress toward Quine’s
goal” (Healey 2017a, b, p. 236).
10 Agents in Healey’s Pragmatist Quantum Theory: A Comparison. . . 235

see, one of the main reasons of tension between agent physicalism and Healey’s
pragmatist approach to quantum mechanics is that, lacking a precise characteriza-
tion of agents that Healey does not provide, we don’t know what PQT is about. To
the extent that agent physicalism requires an exact notion of agent, PQT is incom-
patible with it and quantum theory cannot give a quantum description of an agent. In
this sense, this particular form of physicalism is therefore incompatible with PQT.4
I will begin by briefly reviewing PQT by putting it into a more general
philosophical frame. In the second section I will discuss Healey’s self-proclaimed
quantum realism by arguing that, aside from an implicit endorsement of a contextual
entity realism,5 his position is really instrumentalist despite his claims to the
contrary Healey 2018a. This conclusion will be important to clarify in more details
the ontological status of Healey’s agents and the related issue of agent physicalism.
In the third section, I will discuss the problem of providing some criterion to
distinguish between agents that are quantum systems from agents that are not
and cannot be (like electrons or quarks). As we will argue later, the fact that
Healey’s characterization of agents plausibly relies on the irreducibly epistemic
notion of information is a first argument in favor of the incompatibility between
agent physicalism and PQT.6 This argument will allow me to compare Healey
view of quantum theory with Pitowsky’s approach to the measurement problem,
by discussing in particular the role of observers and the notion of information in
their philosophy.
Before beginning, let me put forth two points to be kept in mind in what follows.
First, in order to be as faithful as possible to Healey’s claims and avoid possible
misunderstandings on my part, it will be necessary to resort to a few quotations from
his texts. The second point is terminological but important nevertheless: “realism
about the wave function
” does not entail a commitment to the existence of a
complex-valued function defined on an abstract mathematical space but entails the
view that
denotes a real physical state that, following many authors and Healey
himself, I will call “the quantum state”.

10.2 A Brief Review of PQT7

It is difficult to summarize an approach to quantum theory as complex as Healey’s

in a few words, also because it has been presented in various papers from different
angles. Here I will just touch on three points that are closely connected to the main
topic of the paper and that are common to all the published versions of PQT, namely

4 The question is whether the failure of agent physicalism also entails the failure of supervenience
physicalism will not be an object of this paper.
5 In PQT, also the reality of quantum particles is a contextual matter.
6 Of course, agents can be characterized in terms of supervenience physicalism.
7 See also Glick (2018).
236 M. Dorato

(P1) Healey’s endorsement of Brandom’s inferentialist theory of meaning,

(P2) His view of the task of the philosopher of quantum theory, and its problematic
reliance on non-quantum magnitude claims (Healey 2012a, p. 740, 2015, p. 11)
or “canonical magnitude claims” as they are later defined (Healey 2017a, b, 80)
and finally,
(P3), the relational nature of quantum states ascriptions, due to the different
perspectives of “physically situated agents” using the Born rule, an expression
often repeated in his texts but that needs some clarification.
P1. In all his presentations, Healey explicitly refers to Brandom’s inferentialist
theory of concepts (Brandom 2000) as a general philosophical background against
which PQT is constructed. There are two aspects that in our context are worth
stressing.
P1.1 The first is that in Brandom’s view, the meaning of any statement (including
those appearing in the formulation of scientific theories) is given by its use, and
in particular by the inferences that it entitles to draw, and not by its represen-
tational capacity. As we will see however – and unlike Suárez more general
inferentialist, non-representationalist approach (2004) – Healey applies his anti-
representationalism explicitly only to mathematical models employed by quantum
theory, without raising the question whether this claim can be extended also to
classical physics. He seems to be committed to the view that all physical models
are non-representational: otherwise he would have to explain how to demarcate
classical mathematical models from quantum mathematical models and what it is
about the former that make it different from the latter.
P1.2 The second aspect is that, beyond Brandom’s philosophical influence,8
the reasons to embrace anti-representationalism are internal to the philosophy
of quantum mechanics. Therefore, if successful, PQT gives an independent and
important support to the general pragmatist agenda proposed by Brandom.9
It is for this reason that PQT might have decisive consequences not just for
the philosophy of quantum mechanics regarded as a fundamental theory of the
physical world, but also for general epistemology and metaphysics. As far as the
former is concerned, PQT invites us to abandon a spectator theory of knowledge, in
which knowing essentially means representing and not intervening in the physical
world (Hacking 1983). Considering that our brain evolved as a means to interact
successfully with the natural world, “agents” and their “knowing how” are to be
regarded as essential components of any kind of naturalistic metaphysics aiming at
understanding the place of human beings in the vast scheme of things.
P2. The second point follows from the first: all physicists know how to use
the theory, but there is no agreement among the different philosophical schools
working on foundations of quantum theory. Philosophers of quantum mechanics

8 Healey’s former student Price (2011) has certainly played a role too.
9 A recent, rather original approach to quantum theory that for various aspects is similar to Healey’s

but more inspired by Wittgenstein is Friederich’s (2015).

10 Agents in Healey’s Pragmatist Quantum Theory: A Comparison. . . 237

have typically assumed that their main task is to answer the question: “what can
the world be like if quantum theory is true?”. Thirty years ago, Healey himself
adopted this viewpoint, since the words in scare quotes are taken from his older
book on quantum mechanics (Healey 1989, p. 6). As he sees it now, however,
the problem with the above question is that it is responsible for a long-standing
dissent in the philosophy of quantum mechanics, which should be put to an end.
The question of interpreting the formalism of the theory in a realistic framework
typically presupposes the so-called “beables”, whose postulation is needed as an
attempt to answer what I take to be an inescapable question: “what is quantum
theory about?” For scientific realist, a refusal to answer this question, is equivalent
to betraying the aim of physics.10
However, according to Healey, in the philosophy of quantum mechanics this is
the wrong question to ask, since it leads to the formulation of theories (Bohm’s,
GRW’s, and many worlds’) underdetermined by their key parameters and by the
evidence.11 For instance, a instrumentalistic view of the quantum state and Bohm’s
realism about the latter state in principle cannot make any predictive difference. This
point has been repeatedly insisted upon also by Pitowsky: “I think that Weinberg’s
point and also Bohm’s theory are justified only to the extent that they yield new
discoveries in physics (as Weinberg certainly hoped). So far they haven’t” (Pitowsky
2005, p. 4, my emphasis).12
The new task that Healey assigns to PQT is not to explain the success of the Born
rules by appealing to some constructive theory in the sense of Einstein (1919), but
rather to explain how quantum theory can be so empirically successful even refusing
to grant both representational power to its models and “reality” to the quantum
state and the Born’s rule, in a sense of “reality” to be carefully specified below
(Healey 2018a).13 Somewhat surprisingly, he claims that his pragmatist approach,
somehow strongly suggested by the theory itself, constitutes the new revolutionary
aspect of quantum mechanics (2012a, b, 2017b, p. 121). This claim should strike
the reader as particularly controversial for at least two reasons. Firstly, as this
section illustrates, a pragmatistic philosophy is at least in part presupposed by
Healey and, like any philosophical approach to QT, cannot be read just from the
data and the theory. Secondly, despite some differences that basically involve the
idea of complementarity, Healey’s PQT rather than being so new, is rather close to
Bohr’s philosophy of quantum mechanics, especially if the latter is purified from
myths surrounding the so-called “Copenhagen interpretation” (Howard 2004) and

10 “What is the theory is about” presupposes simpleminded answers like: classical mechanics is
about particles acted upon by forces, and classical electromagnetism is about fields generated by
charged particles.
11 For this position, see also Bub (2005).
12 Pitowsky is referring to Weinberg’s attempt to reconstruct gravity on a flat spacetime (ibid.).
13 In particular when discussing a crucial but obscure passage in which he claims that “a quantum

states is not an element of reality (it is not a beable”) since “the wave function does not represent
a physical system or its properties. But is real nonetheless” Healey (2017c, p. 172).
238 M. Dorato

read more charitably in light of more careful historical reconstructions (Faye and
Folse 2017).14
In order to convince ourselves of the proximity of PQT and Bohr’s philosophy,
consider the first, following point. It is true that Healey’s empiricist understanding
of quantum theory, unlike Bohr’s, presupposes decoherence, but Howard (1994) and
Camilleri and Schlosshauer (2008, 2015) have put forward the claim that this notion
was in some sense already implicit in Bohr’s philosophy of quantum mechanics,
despite the lack at his time of the relevant physical knowledge. Secondly, even if this
somewhat ahistorical claim were rejected, one must consider PQT’s treatment of the
classical-quantum separation. Without a clear criterion that separates the quantum
from the classical regime, the set of magnitudes that Healey refers to as “non-
quantum” can be typically regarded as those described by classical physics. Healey,
however, explicitly denies that “non-quantum” refers to “classical” since, he claims,
it also refers to the property of any ordinary physical object”. However, the term
‘ordinary object’ is rather vague, unless it is intended to refer to objects with which
we interact in our ordinary life and in our ordinary environment. Consequently, the
reports of experimental results via what he calls NQMC (non-quantum magnitudes
claims)15 involve objects that can be described only by classical physics. This
claim resembles Bohr’s appeal to classical physics as the condition of possibility to
describe the quantum world: “indeed, if quantum theory itself provides no resources
for describing or representing physical reality, then all present and future ways of
describing or representing it will be non-quantum” (Healey 2012a, p. 740).
In addition, there are two close analogies between PQT and Bohr’s empiricist
philosophy of quantum mechanics. The first involves Bohr’s pragmatism, whose
importance for his philosophy has been supported by highly plausible historical
evidence (see among others Murdoch 1987 and Folse 1985). Furthermore, in force
of his holistic view of quantum mechanics, Bohr – analogously to Healey, −
would not deny that there is (a weak) sense in which a quantum state refers to or
describes genuinely physical, mind-independent probabilistic correlations between
measurement apparata and quantum systems. If Healey denied any representation
force to the wave function, he would be in the difficult position to justify the claim
that only classical models, unlike quantum models, represent or describe physical
entities. Otherwise, whence this epistemic separation if one cannot make recourse
to an ontological distinction (remember that in PQT there are no beables)?
These remarks, which are put forward to show the persisting though often
unrecognized influence of Bohr’s philosophy of quantum mechanics,16 are not

14 For a neo Copenhagen interpretation that here I cannot discuss, see Brukner (2017).
15 NQMCs (or canonical magnitude claims) are of the form “the value of dynamical variable M on
physical system S lies in a certain interval ”.
16 After the 1935 confrontation with Einstein, physicists attributed to Bohr a definite victory over

Einstein’s criticism. But since the late 1960s surge of interest in the foundations of physics caused
by Bell’s theorem and his sympathy for alternative formulations of quantum mechanics, Bohr has
become to be regarded as responsible – and not just by philosophers – for having “brainwashed
a whole generation of physicists into thinking that the job was done 50 years ago” (Gell-Mann
10 Agents in Healey’s Pragmatist Quantum Theory: A Comparison. . . 239

intended to deny that PQT diverges from the latter and that therefore significantly
enriches the contemporary debate in the philosophy of quantum mechanics.
P3. The last notable aspect, the relationality of quantum systems, touches more
on the metaphysical aspects of PQT, and if even if it was more stressed in the
earlier version of the formulation of PQT, is the most relevant for my purposes
and in any case interesting for its own sake. In order to summarize its main
thrust, it will be helpful to compare it with Rovelli’s relational quantum mechanics
(Rovelli 1996; Rovelli and Smerlak 2007; Laudisa and Rovelli 2008; Dorato 2016).
Formally, PQT does not accept the eigenstate-eigenvalue link. Philosophically, the
difference involves Healey’s stress on a pragmatist aspect. According to Rovelli’s
view, any interaction of a system with another system (not necessarily an agent)
yields definite results (events). In PQT ascriptions of quantum states are not
relative “to any other system, but only to physically instantiated agents, whether
conscious or not” (Healey 2012a, p. 750, emphasis added). Healey is aware that
his relationism may conflict with agent physicalism but claims to have a way
out: “Since agents are themselves in the physical world, in principle one agent
may apply a model to another agent regarded as part of the physical world. This
second agent then becomes the target of the first, who wields the model. I say
“in principle” since human agents are so complex and so “mixed up” with their
physical environment that to effectively model a human (or a cat!) in quantum theory
is totally impracticable” (Healey 2017b, p. 10). In short, his essential proposal,
to be discussed in what follows, is that “all agents are quantum systems with
respect to other users, but not all quantum systems are agents” (Healey 2012a, b,
p. 750).17 Except for the relational aspect, this statement is analogous to the claim
that every mental event is a physical event, but not all physical events are mental.
Independently or the relational aspect of the claim, the quotation raises another
crucial question to be discussed below: how do we distinguish between quantum
systems that are agents and those that are non-agents? Notice that this question
is essential to any philosophical view that stresses the role of agents in using the
quantum formalism.

10.3 Is PQT Really Realist?

I conclude my brief sketch of PQT by assessing its implications for the issue of
scientific realism which is importanto also for (Healey 2018a). It might be objected
that this is a desperate or useless enterprise, for at least two reasons. Firstly,

1976, p. 29). It is desirable to have a more balanced view of his contribution to the philosophy of
quantum mechanics.
17 While here I will focus on this relational version of PQT as expressed in the 2017’s quotation,

but I will show why the same questions would be relevant also to later non-relational versions of
PQT.
240 M. Dorato

each philosopher has her brand of realism and following Fine (1984) and others,
we might legitimately have the temptation to drop the whole issue in favor of a
natural ontological attitude. Secondly, it is not easy to locate Healey’s position
in the continuous spectrum that goes from full-blown scientific realism to radical
instrumentalism, since he is very careful to stir away from both positions. But for
my purpose however, attempts to solve this problem is important to find out whether
some kind of scientific realism about quantum mechanics is necessary for agent
physicalism.
Here we should distinguish between entity realism in Hacking’s “interventionist
sense” (1983) and any other sort of scientific realism. As far as the former is
concerned, there is no doubt about the fact that Healey is a scientific realist: PQT
describes the unobservable because entities (like quarks or muons) that agents apply
quantum theory to exist independently of them. However, the following six points
ought to convince the reader that, despite Healey’s claim to the contrary (2017a, p.
253 (Healey 2018a)), PQT is a typical instrumentalist position in any other possible
sense of “scientific realism”.
1. Healey claims that his PQT is not outright instrumentalist because of the
normative role of the quantum state, which is justified by the fact that its use
yields correct, empirically confirmed, mind-independent probabilistic predictions.
His “realism” about quantum mechanics could at most imply that the quantum state
can be taken to “refer” or “represent” mind-independent, statistical correlations.
However, statistical correlations about what? Since Healey does not explicitly raise
this crucial question, we may want to consider two possible interpretations. The first
is to read these correlations as holding between observations relating what he calls
“backing conditions” (inputs) via “advice conditions” to outputs or predictions that
must be formulated in the language of classical physics via NQMC (see note 16).
The idea that a theory connects mind-independent empirical correlations between
observations and predictions is one of the features of instrumentalism, so that this
first reading leads to antirealism. In referring to those conditions, the second reading
omits any explicit reference to observations. However, a distinctive feature of PQT
is that these empirical correlations are not to be explained bottom-up by postulating
some beables. On the contrary, it is plausible to attribute Healey the view that
they ought to be regarded as universally valid, empirical regularities like those that
one encounters in thermodynamics or, more generally, in the so-called “theories of
principle” (Einstein 1919). The idea that quantum mechanics ought to be regarded
as a theory of principle in this sense expresses an empiricist philosophical attitude
that is shared also by Bub and Pitowsky (2010). This attitude, however, is not a mark
of a realist approach to physical theories, at least to the extent that such an approach
tries to build constructive theories to explain those empirical regularities.
2. As in all pragmatist theories of truth, there is no attempt to explain the
effectiveness of the Born rule by the claim that quantum theory is approximately
true, since in the framework of a pragmatistic theory there is nothing to explain!
Antirealist philosophers are suspicious of explanations of effectiveness, since if
the quantum algorithm were not effective, it would not be empirically adequate
(a typical refrain of van Fraassen). To the extent that this effectiveness is a brute
10 Agents in Healey’s Pragmatist Quantum Theory: A Comparison. . . 241

fact, truth as correspondence is either a superfluous honorific title or a purely

metaphorical reference to a “mirroring” of models. In this respect, Healey follows
Price in defending minimalist theories of truth (see Price 2011). However, the
legitimate refusal to endorse an inference to the best explanation (van Fraassen
1985, p. 252) based on a more robust representational power of quantum models
is a mark of empiricist/antirealist philosophies of science, whether quantum theory
is involved or not (see Psillos 1999). If one grants that PQT explains in the sense of
Healey (2015), as it does, “more realistic” approaches to quantum theory become
useless or redundant.
3. The third argument in favor of the claim that PQT is deep down instrumental-
istic and involves Healey’s partially spelled out treatment of quantum probabilities.
In order to make it more explicit, it is first useful to distinguish between the
inter-subjective validity (objectivity1 ) and mind independence (objectivity2 ) of
probability claims, and then note that Healey explicitly rejects Qbism’s subjective
views of probability. In this sense he is distant from Pitowsky’s subjectivist account
of probability (Bub and Pitowsky 2010). Independently of what probability and
information correspond to in the physical world, Healey holds that by using the
information (nota bene, an epistemic term) contained in the quantum state, any
physically situated agents would assign the same probability to any measurement
outcome. Given the reliance on information, it necessarily follows that in PQT
probabilities are to be regarded as inter-subjectively valid (i.e. objective1 ) but
still epistemic. The fact that “quantum states are objective informational bridges”
(Healey 2017a) reinforces my point: I know of no approach to ‘information’ that
accounts for it in mind-independent terms:
However, there is a sense in which according to Healey these probabilities are
also mind-independent (objective2 ): “there are real patterns of statistical correlation
in the physical world. Correctly assigned quantum states reliably track these through
legitimate applications of the Born rule” (2017c, p. 172, my emphasis) and the
Born rules “offer a correct or “good advice” to agents whose aim is predicting,
explaining and controlling natural phenomena” (Healey 2017b, p. 102). In Healey’s
philosophy of quantum probability, however, there is an oscillation between these
two senses of “objective”, which are never clearly distinguished. We just read
that, according to Healey, the Born rules do capture objective2 , mind-independent
statistical correlations, a claim that partially explains why they offer “good advice”
and are effective for predictions. On the other hand, ‘to track these correlations”
for him cannot be equivalent to ‘represent’, because the Born rules do not describe.
One reason why they don’t is that “the frequencies they display are not in exact
conformity to that rule—the unexpected does happen...” (2017c, p. 172). It follows
that the expression ‘to track correctly’ can only refer to an unexplained link between
the predictive success ensured by the rules and mind-independent, probabilistic
patterns. The link is unexplained because the Born rules are not derived from
some deeper posit of the theory (as in Bohmian mechanics’ initial distributions
of particles, for example). Healey’s is a legitimate empiricist position but, lacking
a more exact account of the notion of probability while denying representation
force to Born rule, it amounts to a form of probabilistic instrumentalism. Healey
242 M. Dorato

could rebut that postulating objective physical probabilities is not incompatible per
se with the claim that these are not reducible or supervenient upon propensities,
frequencies or Lewisian chances. However, these are the most often discussed
objective accounts of probabilities in the market and postulating objective but
“theoretical probabilities” (a theoretical entity that, according to Healey’s anti-
representationalism, is not “represented” but tracked by the Born rules) is certainly
against a pragmatist spirit. Therefore, until we are told what these objective
probabilities are and what “track” exactly means, we be entitled to conclude that
in PQT the Born rules are just an effective bookkeeping device, given that the
assignment of the quantum state depends entirely on the “algorithm provided by the
Born Rule for generating quantum probabilities” (2012a, b, p. 731). In a word, given
the probabilistic nature of quantum theory, it is safe to assume that Healey’s anti-
representationalism about the Born rule (and the quantum state) essentially depends
on his philosophy of probability, and therefore on the fact that PQT’s probabilities
are objective1 (or intersubjectively valid for all agents) but they are not objective2 .
The question is whether this epistemic view can be regarded as consistent with agent
physicalism and treating agents as physical systems.
4. Another piece of evidence of Healey’s antirealistic approach is his view
of relations between quantum system: “By assigning a quantum state one does
not represent any physical relation among its subsystems: entanglement is not a
physical relation” (Healey 2017b, p. 248, my emphasis). Now, even if Schrödinger
was wrong in claiming that entanglement is “the characteristic trait of quan-
tum mechanics” (Schrödinger 1935, p. 555), there is no question that, in more
realistically-inclined approaches to quantum theory, entangled states are considered
to be real physical relations between physical systems that have no separate identity.
Ontic structural realism is an example of a philosophical approach that takes
relations between quantum entities as ontically prior (French 2016).
This approach to entanglement creates tension with PQT’s reliance on deco-
herence as a way to solve practically the measurement problem, to explain the
emergence of the classical world and to justify its reliance on the epistemic rock
given by NQMC. In particular, decoherence is needed because quantum theory
does not imply and cannot by itself explain the definiteness of our experience
(2017a, p. 99). To the extent that invoking decoherence does not exclude more
realistic attempts at solving the measurement problem – dynamical collapse models,
Bohmian mechanics, Everett-type interpretations – PQT is more antirealistic than
these positions that grant reality to the quantum state and yet also rely on decoher-
ence. Moreover, decoherence presupposes a spreading of entanglement, which is
typically regarded as a genuine, irreversible physical process. But if “entanglement
is not a physical relation” but a mathematical relation among state vectors (2017a,
p. 248), it is problematic to rely on this relation to claim that our ordinary classical
has an approximately classical appearance.
5. Despite what the author claims about the fact that PQT does not explicitly
mention “observables” or “measurement” (2012b, p. 44) – the typical vocabulary
of instrumentalist positions – the centrality of NQMC in Healey’s philosophical
project is incompatible with the view that his approach to quantum theory is
10 Agents in Healey’s Pragmatist Quantum Theory: A Comparison. . . 243

realist. To begin with, non-quantum physical magnitudes obtained in a lab are

measurements outcomes describable with classical physics. And even if many of
these claims refer to systems that are outside the lab, these magnitudes are “...not
part of the ontology of quantum mechanics itself. To provide authoritative advice on
credences in magnitude claims, quantum mechanics must be supplied with relevant
magnitudes—it does not introduce them as novel ontological posits” (Healey 2017b,
p. 233). To this fifth objection, Healey could reply that we cannot and should not ask
more in the case of quantum theory, since its function is to produce observational
results that, as already stressed by Bohr, must necessarily be expressed in classical
language. In this case, however, he owes us an explanation as to why classical
models do not represent just empirical correlations, but can be used to yield deeper
explanations of the latter in terms of entities or properties that help us to “construct”
them, in the same sense in which statistical mechanics grounds thermodynamical
regularities (Einstein 1919).
His response to this point is contained in the following paragraph, hinting to
the contextual and relational nature of the quantum ontology, an aspect that also
characterizes quantum field theory: “the unknown particles, strings, and fields
would still not be in the ontology of the quantum field theory itself: in that
sense they would still be “classical” rather than quantum, even though no prior
theory of classical physics posited their existence. A contextual ontology could
not provide the ultimate building blocks of the physical world” (2017b, pp. 233–
234). In a word, quantum fields “are not a novel ontological posit of quantum
field theories: they are assumables, not beables (ibid.)”. Two counter-objections to
this passage can be raised. First, “assumable” is closely related to “useful posit”,
or even fictional entity, as atoms were for Mach and Poincaré: as a consequence,
Healey’s brand of entity realism is even weaker than that defended by Hacking
(1983). Secondly, it is not clear why contextuality per se should be incompatible
with an ontological commitment to entities whose properties are not intrinsic or
monadic, but fundamentally relational or dispositional, as they are in Rovelli’s
relational quantum mechanics.18 But PQT’s contextualism concerns entities and
not the metaphysical nature of their properties and in this sense it is weaker than
Rovelli’s.
6. In general, pragmatist approaches to knowledge, and therefore also to physical
theories (PQT included), seem to be particular fit to justify a wide-ranging,
naturalistic worldview. Within naturalism as it usually intended, beliefs in general
(scientific ones included) are typically regarded as more or less effective guides to
action. Implicitly, this applies also to PQT: the reliability of the agents’ beliefs about
the effectiveness of attributing a certain quantum state to physical systems depends
on the past predictive success of this attribution. As the previous section has made
abundantly clear, PQT has a definite empiricist and instrumentalist flavor, witnessed
by the fact that typical empiricist epistemologies rely on conditions of assertibility
for certain hypotheses, and therefore stress the epistemic limitations of knowers –

18 Dorato (2016).
244 M. Dorato

or, as Healey has, of “physically situated agents or users”. Some of these conditions
are obviously dictated by the theory itself.
It seems to me that the many senses in which one can be a realist about
quantum mechanics discussed above are sufficient to refute the claim that PQT is
a non-instrumentalist approach to quantum theory. Insisting that PQT is “realist”
about the quantum state while denying the status of a beable is realism only in
name, since PQT’s “realism” about the wave function amounts to the claim that
assigning a quantum state is an effective instrument to predict the outcomes of an
experiment. As hinted above, this conclusion is not just the product of an exegetic
effort, but will be important in the next section, when I will deal with the question
of agent physicalism.19

10.4 The Incompatibility Between PQT and Agent

Physicalism

Interpretations of quantum theory are doubtlessly controversial, and physicalism

is also controversial and vaguely defined Stoljar (2010), but the current scientific
outlook has it that conscious agents are complex “physical systems”. It is not
plausible to suppose that advances in the neurocognitive sciences may require a
change in the way we interpret quantum theory. However, even without defending
agent physicalism, I take it that it would still be important to establish whether
PQT is compatible with it or not. If agent physicalism were more credible than any
approach to quantum theory that violates it (something that here I will not discuss),
a defense of PQT would become more difficult.
Given that agents are physically situated users, Healey’s empiricist denial of
realism about the wave function and the existence of beables prima facie cannot
be regarded as incompatible with some generic of weak forms of reducibility (or
supervenience) of the mental states of agents to (over) their physical states and
therefore with agent physicalism or the possibility of applying quantum mechanics
to agents. Unfortunately, even granting to Healey that quantum theory does not
provide a fundamental and complete ontology of the physical world on which

19 Infairness to Healey, it is very important to qualify these six points: PQT is not incompatible
with the idea that realism ought to be the goals of scientific theorizing. I think Healey would
agree with the idea that realism, like physicalism, are stances or attitudes (van Fraassen 1985;
Ladyman and Ross 2007), that is, they are theses about the aim of science. One way to express this
desideratum is as follows: “a physical theory ought to be about something and therefore describe
physical reality”. If he were right to argue that quantum theory has not beables of its own, he
would be entitled to be against the view that quantum theory, in its present form at least, has the
resources or instruments to achieve this aim, but he could nevertheless argue that realism about
the wave function as a beable may have an important heuristic role, similar to an idea of reason in
the sense of Kant’s transcendental dialectic. Another possibility is that by endorsing some kind of
“piecemeal realism” (Fine 1991), this conclusion may just apply to posits within quantum theory
and not to quantum theory in general or to other physical theories.
10 Agents in Healey’s Pragmatist Quantum Theory: A Comparison. . . 245

everything else supervenes, I will now show that the attempt to rescue PQT from
the charge of incompatibility with agent physicalism is doomed to fail.
Let us begin by noting that in his original 2012a’s formulation of PQT, Healey
states the following principle:
RPQT any physical system whatsoever (agents included) can be a quantum system
and therefore be assigned a quantum state only relative to an agent applying the
algorithm of quantum theory. Healey 2012a, p. 750)
RPQT is a necessary consequence of any agent-centered approach to quantum
theory. The essentially relational aspect of PQT depends in fact on the meaning
of “agent” and cannot be avoided by the kind of pragmatist approach to quantum
theory defended by Healey. Notice in addition that Healey obviously admits that
not all quantum systems can be regarded as agents: “all agents are quantum systems
with respect to other agents, but not all quantum systems are agents” (Healey 2012a,
p. 750, my emphasis),
Consider the following argument, where “physical” means describable in princi-
ple by current physical theories
1. All quantum systems are physical systems;
2. According to RPQT above, all physical systems (electrons, quarks, atom, agents
etc.) possess a quantum state and are quantum systems not intrinsically, but only
relationally, that is, only with respect to real (or hypothetical) agents who use
information provided by these systems.
3. Some quantum systems are not agents, but all agents are quantum systems only
relative to other agents.
4. PQT does not provide any intrinsic and exact criterion to distinguish between
quantum systems that are agents from those that are not agents.
5. agent physicalism requires an intrinsic and exact physical characterization of the
notion of agent or a physically instantiated user of key notions of the theory.
Conclusion
PQT is incompatible with physicalism
Let us discuss these premises in turn. Premise (1) must be granted by any
non-dualistic approach to quantum theory and is accepted by PQT20 : if it were
false, some quantum objects (agents are the most plausible candidates) would be
non-physical and my claim would be proved. Premise 2 expresses the essential

20 Also, the converse of (1) holds, since it is reasonable to claim that all physical objects are
quantum objects. Notoriously size, which looks the most plausible candidate to be an intrinsic
criterion to separate the two realms, is a non-starter. Microscopic systems like electrons, neutrons or
photons are quantum systems par excellence, but also mesoscopic quantum systems need quantum
mechanics for their description: C60 fullerene molecules, to which also Healey refers, are visible
with electron microscopes. Very recently, scientists in California managed to put an object visible
to the naked eye into a mixed quantum state of moving and not moving (O’Connelly et al. 2010). It
is possible to claim that even stars are quantum systems, given that in their nuclei there are nuclear
reactions.
246 M. Dorato

commitment of PQT in any of its form21 and therefore must be accepted by any
agent-centered view of quantum theory: the recourse to agents or users is necessary
to PQT, otherwise Healey’s approach would not be pragmatist. Premise 3 is a
quotation from Healey’s text (see above) and must be granted: not all physical
systems (say electrons) can apply the algorithm provided by the Born rule, but only
agents can, whatever an agent is.
As far as I can tell, the truth of premise 4 also relies on textual evidence and I
will now show why, as a consequence of the vagueness of the notion of ‘a physically
instantiated agent’ – and despite the fact that needs physical agent for its formulation
(see premise 2) – PQT does not tell us what quantum theory is about.
Relying on a passage from Healey (2017c, p.137), he could rebut this claim by
noting that quantum theory is about the mathematical structures that figure in its
models.22 However, as a proposal to understanding the theory, PQT must of course
refer to the role that agents have in making sense of the functions of the key terms
of the theory (Born rule, quantum state, etc.) by applying them to quantum systems.
Note furthermore that the passage mentioned above justifies the claim that PQT
must refer to agents, not in the simpleminded sense that any physical theory needs
agents for its formulation and implementation. Even in model-theoretical accounts
of scientific theories, a physical theory must be able to interpret its formulas, lest it is
devoid of any physical meaning: classical mechanics is not just about the family of
its mathematical models. And given the refusal to give a representationalist account
of the models of quantum theory, PQT relationalism as expressed by 2 implies that
PQT is also about agents relationally regarded as special quantum systems.
Now the problem is that in order to distinguish in an exact way quantum systems
that are agents from those that are not we need an intrinsic criterion, which Healey,
unfortunately, does not provide: “Agents are generic, physically situated systems
in interaction with other physical systems: . . . the term ‘agent’ is being used very
broadly so as to apply to any physically instantiated user of quantum theory, whether
human, merely conscious, or neither” (2012a, p. 752, my emphasis). However,
we do have an intuitive idea of what agents are: they may be either conscious or
unconscious robots acting like conscious agents (as Healey himself has it in the
same passage, 2012a, p. 752), but atoms, electrons, photons and quarks are certainly
not agents in any sense of the word.
Once again, if we want to know what the theory is about, the distinction between
quantum systems that are agent and quantum systems that are not must be given
in intrinsic terms and must be exact in Bell’s sense (see note 19). The point of
the argument is that the possibility of formulating agent physicalism needs an
exact formulation of the notion a physically instantiated user of the mathematical
apparatus of the theory that, due to the irremediable vagueness of “agent”, PQT is
not able to give. It follows that, in its current form at least, PQT is incompatible
with agent physicalism and this conclusion holds independently of the claim that

21 As a response to a criticism, this must hold also in later, less relational versions of 2.
22 This quotation has been reminded to me by one of the referees.
10 Agents in Healey’s Pragmatist Quantum Theory: A Comparison. . . 247

quantum theory provides a fundamental ontology of the physical world (realism

about the wave function), a view of physicalism which Healey rejects and that would
be unfair to attribute to him.

10.5 What Is an Agent in PQT?

The conclusion of the previous section raises immediate questions: why should
PQT be saddled with the task to describe physically instantiated agents in quantum
terms? Shouldn’t this problem be left to cognitive scientists? “Quantum theory, like
all scientific theories, was developed by (human) agents for the use of agents (not
necessarily human: while insisting that any agent be physically situated, I direct
further inquiry on the constitution of agents to cognitive scientists”) (Healey 2017a,
p. 137). In view of this, don’t alternative formulations or approaches to quantum
theory also suffer from the same problem? Of course, if supervenience physicalism
holds in these latter approaches, it also holds in PQT, since in PQT agents are
quantum systems, even though relatively to other agents.
However, the problem is that, unlike what happens in Everettian, Bohmian
and dynamical-collapse quantum mechanics – PQT cannot account for agents in
quantum terms (and therefore in physical terms) even in principle. In Bohmian
mechanics, agents are physical systems that can in principle be described in
quantum terms: there is no difference between quantum systems that are agents and
quantum systems that are not. An analogous conclusion holds in Everettian quantum
mechanics, whose bizarre ontological consequences (many worlds or many minds)
were put forth to avoid considering agents as different in principle from any other
quantum systems. Dynamical collapse models, in the flash version for instance,
regard agents as constituted by galaxies of “flashes” in the same sense in which a
table is. Analogous considerations hold for the matter field version of GRW. The
fact that in PQT agent physicalism fails explains why it is different from these
approaches: agents must be treated in a different way from non-agents but we are not
told in exact physically terms why this should be so. How can we make sense in a
clearer way of the status of agents qua physically instantiated user of a mathematical
model on the one hand and the physical electrons and quarks on the other?
In order to avoid explaining why agents are distinguished from non-agents,
Healey could regard the notion of ‘agent’ as “primitive”. In one sense, “primitive”
as referred to agents would entail that deeper, constructive explanations (possibly
involving beables) of how the physical interactions or information exchange
between agents and quantum systems can occur can be altogether avoided. This
move would also eschew the charge that PQT, which, as we have argued must
essentially be about agents, is silent about the demarcation line separating them
from system that are clearly non-agents: the interaction between agents and non-
agents must be accepted as one that does not require any explanation, since it is
presupposed by an application of PQT by agents.
248 M. Dorato

In this perspective, PQT’s postulated existence of a division line between

agents and non-agents becomes strikingly similar to the well-known demarcation
between the quantum realm and the classical realm advocated by Bohr. As a matter
of fact, PQT failure of agent physicalism suffers from the same problems. The
remarkable similarities between Bohr’s philosophy of quantum theory and PQT,
already sketched in Sect. 2 above, here can be made more explicit.
First of all, Bohr’s approach to quantum theory is as anti-subjectivist as Healey’s
(for this claim about Bohr, see Howard 1994; Dorato 2017). Secondly, and despite
frequent misunderstandings of his position – suggested by the fact that according
to Bohr any experimental report about quantum systems had to be given in the
language of classical physics – in Bohr’s opinion the distinction between the
quantum and the classical realms is to be regarded as contextual and relational
(Zinkernagel 2016). According to Bohr in fact, in certain experimental situations it
is legitimate to treat a classical object like a quantum object: the choice depends
on the whole experimental context.23 Healey refers to the whole experimental
context with the expression “physical situatedness”, but I submit that, despite the
verbal differences, linked to the fact that Bohr’s interpretation does not mention
agents, PQT’s is very similar to Bohr’s. In fact, the relationality of PQT leads us
to claim that an agent A can both apply quantum theory to quantum systems that
are not agents (an electron) and, in other physical contexts, be the target of such
an application by another agent B. These remarks about Healey’s philosophy of
quantum mechanics are not meant to detract from its originality but rather to put the
pragmatist approach in the correct historical light.
While here I cannot argue in more details for this claim, in Bohr’s thought
classical physics has been plausibly likened to the Kantian transcendental condition
of possibility for describing the otherwise unattainable, “noumenal” quantum world
(see MacKinnon 1982; Murdoch 1987): exactly as in PQT, according to Bohr there
cannot be a quantum description of quantum mechanics. Now, while in Healey’s
texts there is no explicit evidence for this Kantian interpretation, agents could be
regarded as those quantum systems that – in virtue of some distinguishing physical
property P that PQT cannot in principle specify in quantum terms – must be
presupposed to regard any physical system as a quantum system. That such a P
is needed depends on the fact that an electron’s properties do not suffice to apply
quantum theory to other quantum systems.
This property P can either physical or epistemic. We have seen that, in virtue
of the failure of agent primitivism, we cannot describe agents in exact quantum
physical terms (we don’t know what distinguishes physically instantiated agents
from non-agents). Consequently, one might be tempted to suppose – against
Healey’s explicit position – that there exist some transcendental epistemic capacity
P in virtue of which agents can play their role as users, a role that electrons cannot

23 See his response to Einstein’s objection, involving the moving screen with the double slit as
reported in Bacciagaluppi and Valentini (2009).
10 Agents in Healey’s Pragmatist Quantum Theory: A Comparison. . . 249

have. By invoking this transcendental reading of PQT, we could give a possible

explanation as of why agents cannot be described in quantum mechanical terms.
However, given the failure of agent physicalism, ‘transcendental condition of
possibility’ would imply ‘epistemic condition of possibility’. In this case, agents’
epistemic conditions of possibility to apply the quantum formalism in PQT would
become primitive. In fact, unlike what happens in other philosophical approaches
to quantum mechanics – any quantum theory that might be invoked to reduce or
explain these states, for its application would presuppose them!24 In a word, to the
extent that agents’ epistemic states play the role of the distinguishing property P
conditions of possibility for applying PQT to the physical world, agents and their
epistemic states would acquire an epistemically primitive status.
If this Kantian reading of the primitive role of agents in quantum theory
were unconvincing because not grounded in the text, one could try to clarify the
implications of PQT by characterizing the property in virtue of which electrons and
quantum agents differ in terms of the notion of information. Agents are primitive
because they have the capacity to manipulate information. In fact, there is an
asymmetry between the acquisition of information on the part of an agent A on a
non-agential quantum system S, and the flow of information going in the opposite
direction: if S is an electron, it cannot acquire information on an agent A except
in a Pickwickian way, at least if “information” is (controversially) regarded as an
irreducibly epistemic term. Once again, the distinction between agents and non-
agents must be regarded as intrinsic, since only agents can be characterized as users
of the information provided by a quantum system, where this information can be
used with the purpose of predicting and explaining its behavior.
The problem is not given by the teleological characters of terms like user’s
‘aims’ or ‘goals’ relative to predictions and explanation of the physical world: Deep
Blue can be attributed the intention of defeating a human chess player. The rub
rather lies is the vagueness of the meaning of “information” (see Timpson 2013).
Healey, however, takes it in “the ordinary sense of the word” (2017c, p. 179),
and therefore in an essentially epistemic sense: it is only an agent that can have
information about something.25 Without a more detailed analysis, information has
no clear physical meaning, despite the fact that it can be measured by Shannon’s
formula. This justifies why it can be regarded as physically primitive: if we cannot
be told in what precise sense and in which way ‘information’ can be considered to
be a relation between two physical systems, we can regard agents manipulating
it as primitive posits of quantum theory. In a word, and similarly to what was
argued before, if information qua epistemic notion is needed to apply quantum
theory to a quantum system, any reducing quantum theory would have to presuppose

24 Thisconclusion is somewhat similar to Frauricher and Renner’s claim that “Quantum theory
cannot consistently describe the use of itself” (2018), an interesting paper which here I cannot
discuss.
25 Compare Bell’s famous questions: “whose information, information about what?”
250 M. Dorato

information, and therefore, in Healey’s understanding of the term, agents’ epistemic

states irreducibly. If a single epistemic state is irreducible, agent physicalism is
false.
Since I leave to the next, final section the task to provide additional arguments
in favor of the view that contemporary information-theoretic approach to quantum
theory are epistemic, I conclude the current section by discussing the following
example, which I take as providing additional arguments to the conclusion that PQT
and agent physicalism are incompatible, and which is a very simplified elaboration
of Wigner’s- type thought experiments (Wigner 1967; Frauchiger and Renner 2018;
Healey 2018a).
Consider a quantum system S that, in Healey’s term, is not a physically situated
agent. Let A be a physically situated agent that extracts (and then uses) information
from S by physically interacting or correlating with it. Suppose that after this inter-
action the experiment reveals “spin up”. Given the relational character of PQT, S has
spin up relatively to agent A, who is certain of this result. At the same time, the joint
system A + S can be regarded as a quantum system by another physically situated
agent B. Relative to B, before her information-gathering physical interaction with the
quantum system A + S, A and S are still in a superposition of states. Consequently,
given the relativity of probability assignments in PQT, it may happen with a certain
probability that after the interaction between the coupled quantum systems A + S
and B, the physically situated agent B reports: “relative to me, A observes spin down
and S has spin down” even if from A’s perspective, A has observed spin up. Suppose
this is exactly what happens after B’s measurement of A + S.
It is very important to notice that, as in Rovelli’s interpretation of quantum
mechanics (1996), this fact does not contradict A’s previous report, since it is simply
a consequence of the relativity of the ascription of quantum states by quantum
agents. Since B can be regarded as a quantum system by A, when B and A correlate,
A will claim “from agent’s B perspective, I have observed spin down even if from
my perspective I have observed spin up” and B will agree with this report.
However, it is one thing to claim that there is no fact of the matter as to what
the spin of the system is independently of the experimental setup and the physical
situatedness of the agents. It is quite another to have to deny the claim that “agent A
has either observed that the spin is up or she has observed that it is down.” Agents’
observations cannot be relative to different interacting systems. There seem to be
only two ways out of this problem. (i) The first is to claim that PQT is incompatible
with the view that measurements have definite results, so that it requires some
kind of Everettian relative-state mending, which however, PQT’s instrumentalism
about the quantum state cannot advocate: remember that in Everettian quantum
mechanics the wave function is all there is. Furthermore, without some many
worlds formulation, we would have two different epistemic states of the same
agents A relative to the same measurement result, which might be regarded as
a failure of supervenience: different conscious perceptions correlate to the same
physical outcomes, so that the former don’t fix the latter. (ii) Quantum theory is
not universally applicable, and in particular not applicable to agents, which must be
regarded intrinsically as classical physical systems having informational access, via
10 Agents in Healey’s Pragmatist Quantum Theory: A Comparison. . . 251

decoherence, to non-quantum magnitude claims. But this way out introduces once
again an unclear distinction between quantum systems and classical systems whose
vagueness the relationality of PQT wanted to avoid.

10.6 Why Pitowsky’s Information-Theoretic Approach

to Quantum Theory Also Needs Agents

In this concluding section, I want to very briefly comment on Healey’s and

Pitowsky’s approaches to the related questions of realism, probability and infor-
mation.

10.6.1 Realism

As to realism, that there has been an evolution in Pitowsky view of quantum theory.
In chapter 6 of Quantum Probability and Quantum Logic (Pitowsky 1989), he seems
to defend a form of underdetermination of theories by data, and his preference for
Bohmian mechanics, were one to choose a hidden variable theory, is clearly stated:
“ . . . Can we produce a hidden variable theory which goes beyond quantum theory
and predicts new phenomena? Psychologically this question seems to be connected
to the problem of realism. It seems that realists are more inclined to consider
hidden variable theories. Logically however, the physical merits of hidden variable
theories are independent of the metaphysical debate. One can take an instrumentalist
position regarding hidden variables and still appreciate their potential application in
predicting new phenomena. Alternatively, one can believe in the existence of hidden
causes, while still holding on to the view that these causes lie beyond our powers of
manipulation. Hidden variable theories may very well lead to new predictions, this
is a matter of empirical judgment. My own bias tends to side with non-local theories
such as Bohm (1952). In this and similarly detailed dynamical theories we may be
able to find a way to manipulate and control the hidden variables and arrive at a new
phenomenological level. In any case, I can conceive of no a-priori argument why
this could not occur” (Pitowsky 1989, p. 181).
While this manipulation is today regarded as out of question, in his earlies
production Pitowsky’s open-mindedness about quantum theory distinguishes his
position from Healey’s more recent work. It is significant, however, that in order
to break the underdetermination, he does not regard the greater explanatory virtue
of Bohm’s theory as sufficient to rule out instrumentalist approaches, as long as
the latter are regarded as having the same epistemic force. The decisive element
would be the production of new experimental evidence, an aspect that today is
unreasonably regarded with skepticism.
In later works, however, Pitowsky has argued that quantum theory can be char-
acterized structurally as a theory about the logic and probability of quantum events,
252 M. Dorato

that the measurement problem is a pseudo-problem, that the quantum state does not
represent reality but just provides an algorithm to assign probabilities. Unlike the
position defended by Healey, according to Pitowsky probabilities are credences that
reflect objective correlations and that are assigned to certain outcomes. As PQT,
also according to Pitowsky decoherence explains the emergence of a classical world
(see, for instance, Bub and Pitowsky 2010, p. 1). These are the typical traits of
instrumentalistic approaches to quantum theory, at least insofar as the quantum state,
denoted by the wavefunction, is regarded as the central notion used by quantum
theory and there is no attempt at giving a dynamic explanation of the definiteness of
the measurement outcomes that does not rely on decoherence. This particular form
of antirealism about the wave function is certainly shared also by Healey, even he
does not explicitly argue that the claim that measurement should not be appear in
the primitive of the theory is the result of “a dogma”.

10.6.2 Probability

The theory of probability defended by the two authors prima facie seems different.
Healey rejects subjectivist view of probability, since, modulo the remarks above,
PQT is described as capable of tracking mind-independent probabilistic correla-
tions. According to Bub and Pitowsky, the quantum state is credence function
(2010), and the objective nature of probability is given by the kinematic constraints
represented by the projective geometry of Hilbert space. Yet, also Bub and Pitowsky
argue that angles in such a geometry are associated with “structural probabilistic
constraints on correlations between events”, where these events are to be regarded
as physical, in analogy with events in Minkowski spacetime, also structured by
spatiotemporal relations. Despite the fact that Healey does not explicitly mention
intersubjectively shared credences, also the probabilistic approach of the two
philosophers is strikingly similar.
The real difference between Healey and Pitowsky’s (and Bub’s) philosophies
of quantum theory involves the relational character of PQT, which is called for by
the agent-centered view of Healey’s approach. We have seen why it is the role of
agents that makes PQT into a relational theory and therefore creates troubles vis à
vis agent physicalism. I will now show that there is an important pragmatic, even
if only implicit component also in Pitowsky’s approach, which is a consequence of
the pivotal role that information has in his late production and that pushes toward
a relational approach, involving conscious agents manipulating and transferring
information.

10.6.3 Information

In order to argue in favor of this claim it is important to figure out in what sense
information can be regarded as physical, given that according to Pitowsky and
10 Agents in Healey’s Pragmatist Quantum Theory: A Comparison. . . 253

Bub, quantum theory can be reconstructed as a theory about the transmission and
manipulation of information, “constrained by the possibilities and impossibilities of
information-transfer in our world” (Bub 2005).26 Recall that according to Healey
information has a merely epistemic sense and plays a minor role in his theory.
However, the fact that agents are necessarily presupposed also in information
theoretic approaches to quantum theory is evident from the fact that only users
can manipulate information regarded as a physical property of objects. Think of
questions like: “what is an information source?”, what does it mean “to broadcast
the outcome of a source?” Is the “no-signaling principle” understandable without
the presence of an agent, and therefore in purely physical terms?
Let us agree to begin with that the no cloning principle applies to machines, since
it claims that “there is no cloning machine that can copy the outcome of an arbitrary
information source” (2010, p. 7). The main question therefore involves the notion
of a source of information. A fair coin, that either land head or tail, is a source of
information (i.e., a classical bit) but from the physical viewpoint it is merely a short
cylinder with two different faces. Only agents, whether conscious or unconscious,
can use these physical properties with some goal or intention (toss it decide who
must go shopping for example), but without a use of this property by someone or
something, there is no source of information, even if information has no semantic
sense, as pointed out by Shannon in his original paper.
Signaling is a word that has a univocal, intentional sense: only agents can signal
to other agents by using physical instruments with the purpose of communicating
something, but the existence of a probabilistic correlation between two experimental
outcomes by itself can carry no information. Consider now a photon, which is a
qubit with respect to its polarization. In this sense, it’s purely physical (possibly
dispositional) properties are analogous to that of a coin, despite the fact that a
qubit is associated to “an infinite set of noncommuting two-values observables
that cannot have all their value simultaneously” (Bub 2016, p. 30). The physical,
agent-independent, property of the photon is its elliptical or linear polarization, the
information it carries and the possibility of using it as a source thereof presupposes
the existence of a user, whether conscious or not, with some intention to do
something. In this sense, information theoretic approaches à la Pitowsky and Bub
are much more similar to Healey’s PQT than it might be thought at first sight. This
should be of no surprise, given the shared claim that the wave function has no
representative power and has a pure algorithmic role. There is a sense according
to which instrumentalist philosophies are more anthropocentric than their more
realistically inclined rivals.

26 “quantum mechanics [ought to be regarded] as a theory about the representation and manipula-
tion of information constrained by the possibilities and impossibilities of information-transfer in
our world (a fundamental change in the aim of physics), rather than a theory about the behavior of
non-classical waves and particles” (Bub 2005, p. 542).
254 M. Dorato

Acknowledgments I thank two anonymous referees for their very precious help on previous
version of this manuscript. In particular, I owe many remarks that I added to the second version of
the paper to the second referee.

References

Bacciagaluppi, G., & Valentini, A. (2009). Quantum theory at the crossroad. Cambrigde: Cam-
brigde University Press.
Bell, J. (1989). Towards an exact quantum mechanics. In S. Deser & R. J. Finkelstein (Eds.),
Themes in contemporary physics II. Essays in honor of Julian Schwinger’s 70th birthday.
Singapore: World Scientific.
Bohm, D. (1952). A suggested Interpretation of the quantum theory in terms of ‘Hidden’ variables,
I and II. Physical Review, 85(2), 166–193. https://doi.org/10.1103/PhysRev.85.166.
Brandom, R. (2000). Articulating reasons: An introduction to Inferentialism. Cambridge, MA:
Harvard University Press.
Breuer, T. (1993). The impossibility of accurate state self-measurements. Philosophy of Science,
62, 197–214.
Brukner, Č. (2017). On the quantum measurement problem. In R. Bertlmann & A. Zeilinger (Eds.),
Quantum (un)speakables II (pp. 95–117). Cham: Springer.
Bub, J. (2005). Quantum theory is about quantum information. Foundations of Physics, 35(4),
541–560.
Bub, J. (2016). Banana world. Quantum mechanics for primates. Oxford: Oxford University Press.
Bub, J., & Pitowsky, I. (2010). Two dogmas about quantum mechanics. In S. Saunders, J. Barrett,
A. Kent, & D. Wallace (Eds.), Many worlds?: Everett, quantum theory & reality (pp. 433–459).
Oxford: Oxford University Press.
Camilleri, K., & Schlosshauer, M. (2008). The quantum-to-classical transition: Bohr’s doctrine of
classical concepts, emergent classicality, and decoherence. arXiv:0804.1609v1.
Camilleri, K., & Schlosshauer, M. (2015). Niels Bohr as philosopher of experiment: Does
decoherence theory challenge Bohr’s doctrine of classical concepts? Studies in History and
Philosophy of Modern Physics, 49, 73–83.
Churchland, P., & Hooker, C. (Eds.). (1985). Images of science: Essays on realism and empiricism
(with a reply from bas C. van Fraassen). Chicago: University of Chicago Press.
Dalla Chiara, M. L. (1977). Logical self-reference, set theoretical paradoxes and the measurement
problem in quantum mechanics. Journal of Philosophical Logic, 6, 331–347.
Dorato, M. (2016). Rovelli’s relation quantum mechanics, antimonism and quantum becoming. In
A. Marmodoro & D. Yates (Eds.), The metaphysics of relations (pp. 235–261). Oxford: Oxford
University Press.
Dorato, M. (2017). Bohr’s relational holism and the classical-quantum interaction. In H. Folse &
J. Faye (Eds.), Niels Bohr and philosophy of physics: Twenty first century perspectives (pp.
133–154). London: Bloomsbury Publishing.
Dürr, D., Goldstein, S., Tumulka, R., & Zanghì, N. Bohmian mechanics. https://arxiv.org/abs/
0903.2601v1
Einstein, A. (1919, November 28). Time, space, and gravitation. Times (London), pp. 13–14.
Faye, J., & Folse, H. (Eds.). (2017). Niels Bohr and the philosophy of physics, twenty-first-century
perspectives. London: Bloomsbury Academic.
Fine, A. (1984). The natural ontological attitude. In J. Leplin (Ed.), Scientific realism. Berkeley:
University of California Press.
Fine, A. (1991). Piecemeal realism. Philosophical Studies, 61, 79–96.
Folse, H. (1985). The philosophy of Niels Bohr. The framework of complementarity. Amsterdam:
North Holland.
10 Agents in Healey’s Pragmatist Quantum Theory: A Comparison. . . 255

Folse, H. (2017). Complementarity and pragmatic epistemology. In J. Faye & H. Folse (Eds.), Niels
Bohr and the philosophy of physics. A-twenty-first-century perspective (pp. 91–114). London:
Bloomsbury.
French, S. (2016). The structure of the world. Oxford: Oxford University Press.
Frauchiger, D., & Renner, R. (2018). Quantum theory cannot consistently describe the use of
itself. Nature Communications, 9, Article number: 3711. https://doi.org/10.1038/s41467-018-
05739-8.
Friederich, S. (2015). Interpreting quantum theory. A therapeutic approach. New York: Palgrave
Macmillan.
Gell-Mann, M. (1976). What are the building blocks of matter? In H. Douglas & O. Prewitt (Eds.),
The nature of the physical universe: Nobel conference (pp. 27–45). New York: Wiley.
Glick, D. (2018). Review of the quantum revolution in philosophy by Richard Healey. philsci-
archive.pitt.edu/14272/2/healey_review.pdf
Godfrey-Smith, P. (2003). Theory and reality. Chicago: University of Chicago Press.
Hacking, I. (1983). Representing and intervening. Cambridge: Cambridge University Press.
Healey, R. (1989). The philosophy of quantum mechanics. Cambridge: Cambridge University
Press.
Healey, R. (2012a). Quantum theory: A pragmatist approach. British Journal for the Philosophy of
Science, 63, 729–771.
Healey R. (2012b). Quantum decoherence in a pragmatist view. Resolving the measurement
problem. arXiv.org > quant-ph > arXiv.
Healey, R. (2015). How quantum theory helps us explain. British Journal for Philosophy of
Science, 66, 1–43.
Healey R. (2017a). Quantum-Bayesian and pragmatist views of quantum theory. In E. N.
Zalta (Ed.), The Stanford encyclopedia of philosophy (Spring 201a edition). URL: https://
plato.stanford.edu/archives/spr2017/entries/quantum-bayesian/
Healey, R. (2017b). The quantum revolution in philosophy. Oxford: Oxford University Press.
Healey, R. (2017c). Quantum states as objective informational bridges. Foundations of Physics, 47,
161–173.
Healey, R. (2018a, unpublished manuscript). Pragmatist quantum realism. British Journal for
Philosophy of Science.
Healey, R. (2018b). Quantum theory and the limits of objectivity. Foundations of Physics, 48(11),
1568–1589.
Howard, D. (1994). What makes a classical concept classical? Toward a reconstruction of Niels
Bohr’s philosophy of physics. In Niels Bohr and contemporary philosophy (Vol. 158 of Boston
studies in the philosophy of science) (pp. 201–229). Dordrecht: Kluwer.
Howard, D. (2004). Who invented the ‘Copenhagen interpretation?’ A study in mythology.
Philosophy of Science, 71, 669–682.
Ladyman, J., & Ross, S. (2007). Everything must go. Metaphysics naturalized. Oxford: Oxford
University Press.
Laudisa, F., & Rovelli, C. (2008). Relational quantum mechanics. In E. N. Zalta (Ed.), The
Stanford encyclopedia of philosophy (Fall 2008 edition). URL: http://plato.stanford.edu/
archives/fall2008/entries/qm-relational/
MacKinnon, E. (1982). Scientific explanation and atomic physics. Chicago: Chicago University
Press.
Maudlin, T. (2007). The metaphysics within physics. Oxford: Oxford University Press.
McGinn, C. (1993). Problems in philosophy: The limits of inquiry. Cambridge: Basil Blackwell.
McLaughlin, B., & Bennett, K. (2014). Supervenience. In E. N. Zalta (Ed.), The Stanford
encyclopedia of philosophy (Spring 2014 edition). URL: https://plato.stanfrd.edu/archives/
spr2014/entries/supervenience/
Minkowski, H. (1952). Space and time. In H. A. Lorentz, A. Einstein, H. Minkowski, & H. Weyl
(Eds.), The principle of relativity: A collection of original memoirs on the special and general
theory of relativity (pp. 75–91). New York: Dover.
Murdoch, R. (1987). Niels Bohr’s philosophy of physics. Cambridge: Cambridge University Press.
256 M. Dorato

O’Connell, A. D., Hofheinz, M., Ansmann, M., Bialczak, R. C., Lucero, E., Neeley, M., Sank, D.,
Wang, H., Weides, M., Wenner, J., Martinis, J. M., & Cleland, A. N. (2010). Quantum ground
state and single-phonon control of a mechanical resonator. Nature, 464, 697–703.
Pitowsky, I. (1989). Quantum probability, quantum logic. Berlin: Springer.
Pitowsky, I. (2005). Quantum mechanics as a theory of probability. arXiv:quant-ph/0510095v1.
Price, H. (2011). Naturalism without mirrors. Oxford: Oxford University Press.
Psillos, S. (1999). How science tracks truth. London/New York: Routledge.
Quine, W. V. (1981). Theories and things. New York: The Belknap Press.
Rovelli, C. (1996). Relational quantum mechanics. International Journal of Theoretical Physics,
35, 1637–1678.
Rovelli, C., & Smerlak, M. (2007). Relational EPR. Foundations of Physics, 37, 427–445.
Schrödinger, E. (1935). Discussion of probability relations between separated systems. Proceed-
ings of the Cambridge Philosophical Society, 31, 555–563; 32, 446–451.
Stoljar, D. (2010). Physicalism. New York: Routledge.
Stoljar, D. (2017). Physicalism. The Stanford Encyclopedia of Philosophy (Winter 2017 Edition)
(E. N. Zalta (Ed.)). https://plato.stanford.edu/archives/win2017/entries/physicalism/
Suárez, M. (2004). An inferential conception of scientific representation. Philosophy of Science,
71(Supplement), S767–S779.
Timpson, C. (2013). Quantum information theory. Oxford: Clarendon Press.
Van Fraassen, B. (1985). Empiricism in the philosophy of science. In P. Churchland & C.
Hooker (Eds.), Images of science: Essays on realism and empiricism (pp. 245–308). Chicago:
University of Chicago Press.
Van Fraassen, B. (2002). The empirical stance. Princeton: Princeton University Press.
Wigner, E. P. (1967). Remarks on the mind–body question. In Symmetries and reflections (pp.
171–184). Bloomington: Indiana University Press.
Zinkernagel, H. (2016). Niels Bohr on the wave function and the classical/quantum divide. Studies
in History and Philosophy of Modern Physics, 53, 9–19.
Chapter 11
Quantum Mechanics As a Theory of
Observables and States (And, Thereby,
As a Theory of Probability)

John Earman and Laura Ruetsche

Abstract Itamar Pitowsky contends that quantum states are derived entities,
bookkeeping devices for quantum probabilities, which he understands to reflect the
odds rational agents would accept on the outcomes of quantum gambles. On his
view, quantum probability is subjective, and so are quantum states. We disagree. We
take quantum states, and the probabilities they encode, to be objective matters of
physics. Our disagreement has both technical and conceptual aspects. We advocate
an interpretation of Gleason’s theorem and its generalizations more nuanced—and
less directly supportive of subjectivism—than Itamar’s. And we contend that taking
quantum states to be physical makes available explanatory resources unavailable
to subjectivists, explanatory resources that help make sense of quantum state
preparation.

Keywords Probability · Quantum theory · Subjectivism · State preparation ·

Pitowsky · Quantum field theory

11.1 Introduction

Our title pays homage to Itamar Pitowsky’s brilliant “Quantum Mechanics As a

Theory of Probability” (Pitowsky 2006), and at the same time it indicates our
(partial) disagreement with the view of the foundations of quantum mechanics

J. Earman
Department of History and Philosophy of Science, University of Pittsburgh, Pittsburgh, PA, USA
e-mail: jearman@pitt.edu
L. Ruetsche ()
Department of Philosophy, University of Michigan, Ann Arbor, MI, USA
e-mail: ruetsche@umich.edu

© Springer Nature Switzerland AG 2020 257

M. Hemmo, O. Shenker (eds.), Quantum, Probability, Logic,
Jerusalem Studies in Philosophy and History of Science,
https://doi.org/10.1007/978-3-030-34316-3_11
258 J. Earman and L. Ruetsche

promoted in this paper.1 We cannot begin to do justice here to the richness of

Itamar’s paper, and we confine our discussion principally to two of Itamar’s central
theses:
T1: The lattice of closed subspaces of Hilbert space is “the structure that represents the
‘elements of reality’ in quantum theory” (p. 214)
T2: “The quantum state is a derived entity, it is a device for bookkeeping quantum
probabilities” (ibid.), where these are understood as the credences ideally rational
agents would assign events whose structure is described by T1.

We agree very much with the spirit of T1, but we emphasize that T1 needs to be
appropriately generalized to cover the branches of quantum theory that go beyond
ordinary non-relativistic QM, which was Itamar’s focus. And we would underscore,
even more strongly than Itamar did, how the structure that quantum theory gives
to the ‘elements of reality’—or as we would prefer to say, to quantum events or
propositions—shapes quantum probabilities.
Our disagreement with Itamar concerns T2, and has technical and conceptual
dimensions. Our fundamental unhappiness is conceptual. We contend that T2
mischaracterizes the nature of quantum probabilities, and we propose inverting T2
to the counter thesis
T2∗: Not only is the quantum state not a derived entity, and so not merely a device for
bookkeeping quantum probabilities, rather the quantum state has ontological primacy in
that quantum probabilities are properties of quantum systems that are derived from the
state.

Although T2∗ gainsays T2’s claim that quantum states are derived entities, the
theses differ on a deeper matter. Taking quantum states to codify objective features
of quantum systems T2∗ implies that quantum probabilities are objective, observer-
independent probabilities—objective chances if you care to use that label. The
territory, between T2∗ and a view asserting quantum states to be derived from
objective probabilities, may not be worth quibbling over. However, Itamar denies
that quantum probabilities are objective. His T2 claims not only that quantum states
are derived entities but also explains why they are derived entities. On Itamar’s
view, the event structure T1 come first, and then come the probabilities, in the
form of credences rational agents would assign events so structured. The states
are bookkeeping devices for these credences. Itamar’s interpretation of quantum
probability is personalist, and an element of a striking and original interpretation
of QM. Itamar’s personalism widens the distance between T2 and T2∗ to include
territory that is worth fighting over.
Let’s briefly explore that territory. Naively viewed, quantum theory describes an
event structure and a set of probability distributions over that event structure. The
sophisticated might question the division of labor this naive picture presupposes.
Which of the elements italicized above does physics supply and which (if any) do
we, qua rational agents, supply? For Itamar, physics supplies only the event structure

1 See also Pitowsky (2003) and Bub and Pitowsky (2010).

11 Quantum Mechanics As a Theory of Observables and States (And. . . 259

T1; the rest is the work of rational agents. Quantum probability distributions reflect
betting odds ideal rational agents would accept on “quantum gambles”; quantum
states are just bookkeeping devices for those probability distributions.2 By contrast,
while we also accept T1, we expect physics to do more work. And so we assert
T2∗: We take it that physics supplies quantum probabilities, which probabilities are
objective, independent of us, and determined by physical quantum states.
On Itamar’s view, rationality and only rationality constrains what quantum proba-
bilities, and so what quantum states, are available. Our view countenances additional
constraints. If probabilities are physical, both physical law and physical contingency
can govern what quantum probabilities are available for physical purposes. It’s clear
that our explanatory resources are richer than Itamar’s. Rationality on its own can’t
explain Kepler’s laws, but Newton’s Law of Universal Gravitation can. And LUG
on its own can’t explain why Halley’s comet appears every 75 years, but celestial
contingencies (in collaboration with physical law) can. What’s not so obvious:
whether there are any explanations essential to making sense of quantum mechanics
to which our resources are adequate but Itamar’s are not.
Enough has been said to introduce and situate our technical disagreement with
T2. The main result Itamar cites in support of T2 is known as Gleason’s theorem.
By itself the original Gleason theorem does not yield T2; by itself the theorem
says nothing about the ontological priority of probability measures over states, nor
does it rule on whether quantum probabilities are personalist or objective. Finally,
Itamar’s claim that the Born Rule follows from Gleason’s theorem rests on a tacit
restriction on admissible probability distributions/physically realizable states.3 As
we will discuss, how one justifies these restrictions depends on one’s interpretation
of quantum probabilities and/or the status assigned to states. Arguably proponents of
T2∗ have better grounds than proponents of T2 for adopting the restrictions, which
do critical work in attempts to make sense of quantum state preparation.
§§2–4 will explore these technical matters in more depth. §5 and beyond
elaborate and adjudicate our conceptual disagreement with Itamar. The standard by
which both views are judged is: do they explain why QM works as well as it does?
That is, do they make sense of our remarkably successful laboratory practice of
preparing systems, attributing them quantum states, extracting statistical predictions
from those states, and performing measurements whose outcomes confirm those
predictions? With respect to this standard, Itamar’s version of T2 holds both promise
and peril. The promise is that it leads to a dissolution of the notorious measurement
problem.4 The peril is that T2 looks shaky in the face of two explanatory demands:
first, the demand to explain the fact that (some) quantum states can be prepared; and,

2 For details of Pitowsky’s account of quantum gambles, see his (2003, Sec. 1).
3A further technical quibble, readily addressed, is that the version Itamar discusses applies only to
the special setting of non-relativistic QM; §4.2 extends his discussion.
4 Pitowsky (2003) and Bub and Pitowsky (2010) call it the “‘big’ measurement problem,” which

they distinguish from the “small” measurement problem. More on this below.
260 J. Earman and L. Ruetsche

second, the demand to explain the fact that the probabilities induced by prepared
states yield the correct statistics for outcomes of experiments.
The first demand is the topic of §§5–8, which describe how each view fields the
challenge of state preparation, both in ordinary QM and in more general settings
(QFT, thermodynamic limit). An attractive account of state preparation appeals to
physical states—which seems to favor T2∗. In fairness, however, we indicate how
the formal account of state preparation offered by quantum theory can be turned into
an alternative account of state preparation as belief state fixation—a repurposing
congenial to advocates of T2. However, the repurposing requires a restriction on the
space of admissible credences that may not be motivated by demands of rationality
alone. We also confess that we are in danger of being hoisted on our own petard.
Because the quantum account of state preparation seems to falter in QFT, our
insistence that T1 be generalized to cover branches of quantum theory beyond
ordinary QM appears to weaken the support consideration of state preparation lends
T2∗. We review some results that promise to circumvent this roadblock.
Turning to the second demand, to make sense of our capacity to complete
measurements confirming quantum statistical predictions, §9 confesses that T2∗
runs smack into the measurement problem. It contends that Itamar’s view also
founders on this explanatory demand. We argue that the measurement problem is
a real problem that needs to be solved rather than dissolved. A concluding section
recaps the main points of our discussion, and relates them to another foundations
question dear to Itamar’s heart: just how much realism about QM is it possible to
have?
We leave it to the reader to render final judgment on the relative merits of T2
vs. T2∗. All we insist upon is that knocking heads over T2 vs. T2∗ is a good way
of sharpening some of the most fundamental issues in the foundations of quantum
theory.
We begin our discussion with a reformulation of thesis T1 designed to cover all
branches of quantum theory.

11.2 Restatement of Thesis 1

We take a quantum system to be characterized by an algebra of observables and a

set of states on the algebra. For the moment we focus on the algebra of observables,
which we take to be a von Neumann algebra N acting on a Hilbert space H.5 The
elements of quantum reality—or, as we would prefer to say, the quantum events or
propositions—are represented by the elements of the projection lattice P(N) of N
(see Appendix 1), and quantum probability is then the study of quantum probability
measures on P(N). This, which we denote by T1∗, is the promised generalization

5 For
the relevant operator algebra theory the reader may consult Bratelli and Robinson (1987) and
Kadison and Ringrose (1987).
11 Quantum Mechanics As a Theory of Observables and States (And. . . 261

of Itamar’s T1. We feel confident that he would endorse it. It is hardly a novel or
controversial thesis; indeed, T1∗ is the orthodoxy among mathematical physicists
who use the algebraic approach to quantum theory (see, for example, Hamhalter
2003).6
A quantum probability measure on P(N) is a map Pr : P(N) → [0, 1] satisfying
the quantum analogs of the classical probability axioms:
(i) Pr(I ) = 1 (I the identity projection)
(ii) Pr(E1 ∨ E2 ) = Pr(E1 ) + Pr(E2 ) whenever projections E1 , E2 ∈ P (N) are mutually
orthogonal.

The requirement (ii) of finite additivity can be strengthened to complete additivity

(ii∗ family {Ea } ∈ P (N) of mutually orthogonal projections, Pr(∨a Ea ) =
) For any
Pr( Ea ) = Pr(Ea ).7
a a

When N acts on a separable H any family of mutually orthogonal projections

is countable and, thus, complete additivity reduces to countable additivity. When
N acts on a finite dimensional H any family of mutually orthogonal projections
contains only a finite number of members and so only finite additivity is in play.
The different additivity requirements will play an important role in what follows.
Different branches of quantum theory posit different von Neumann algebras of
observables and, thus, according to T1∗, different quantum event structures. In the
case of ordinary QM (sans superselection rules) the algebra of observables is taken
to be the Type I factor algebra B(H), the algebra of all bounded operators acting
on H. Here the elements of P(N) are in one-one correspondence with the closed
subspaces of H, in accord with Itamar’s T1. In relativistic QFT the algebras of
observables associated with open bounded regions of spacetime are typically Type
III von Neumann algebras. This requires a significant modification of T1 since
these algebras contain only infinite dimensional projections. In contrast to Type III
algebras, Type II algebras contain finite dimensional projections; and in contrast
to Type I algebras they contain no minimal projections. Although ergodic measure
theory played a role in the construction of the first examples of Type II algebras
(see Petz and Rédei 1996), these algebras haven’t found wide physical application.
Superselection rules provide a more subtle way in which T1 has to be modified. In
the case of non-relativistic QM a superselection rule is modeled by a decomposition
of the Hilbert space into a direct sum H = ⊕j Hj and a decomposition of the algebra
of observables into the direct sum algebra ⊕j B(Hj ). A projection onto a ray of H
that cuts across the superselection sectors of Hilbert space is not in the algebra of
observables.

6 This is not to say that T1∗ is uncontroversial among philosophers of physics. Bohmians would

deny that T1∗ captures the fundamental event structure.

7 The sum of probabilities for a transfinite collection is understood as the sup of sums of finite

subcollections. The sup exists because the sequence of sums over ever larger finite subcollections
gives a non-decreasing bounded sequence of real numbers, and every such sequence has a least
upper bound.
262 J. Earman and L. Ruetsche

The different event structures T1∗ extracts from different von Neumann algebras
naturally bring with them differences in the structure of quantum probabilities. In
general once the algebra of observables N is fixed so are the non-measurable events:
the latter consist of the complement of P(N) in the set of all projections on H. In
ordinary QM there is no analog of the phenomenon of non-measurable events in
classical probability. All projections on H are in the ordinary QM event algebra;
therefore all lie in the range of a probability measure defined on the algebra. In
relativistic QFT, by contrast, there are non-measurable events galore. The Type III
local algebras endemic there contain no finite dimensional projections and, thus,
every finite dimensional projection on a Hilbert space H on which a Type III algebra
acts corresponds to a non-measurable event. Furthermore, the theory offers a prima
facie explanation of why some events are non-measurable: they are not assigned
probabilities because they do not correspond to observables.
Different types of algebras of observables have different implications for the
status of additivity requirements for quantum probabilities. In particular, in ordinary
QM where the algebra of observables is a Type I factor, countable additivity entails
complete additivity unless the dimension of the Hilbert space is as great as the least
measurable cardinal (see Drish 1979 and Eilers and Horst 1975). This entailment
does not hold for the Type III algebras encountered in QFT.
In the following sections we will see that differences in the algebras of observ-
ables also make for differences in the possible states, which in turn imply differences
in the possible probability measures.

11.3 The Relation Between Quantum States and Quantum

Probabilities

11.3.1 Quantum States

Quantum states are expectation value functionals on the observable algebra N—

technically, complex valued, normalized positive linear functionals. A state ω is
said to be mixed (or impure) iff it can be written as a convex linear combination of
distinct states, i.e. there are distinct states φ 1 , φ 2 such that ω = λ1 φ 1 + λ2 φ 2 with
0 < φ 1 , φ 2 < 1 and φ 1 + φ 2 = 1. A vector state ω is a state such that there is
a |ψ ∈ H where ω(A) = ψ|A|ψ for all A ∈ N. Vector states are among the
normal states, which can be given a number of equivalent characterizations. Here
we quote one relevant result:
Theorem 1. The following conditions are equivalent for a state ω on a von Neumann algebra
N acting on H:
 
(a) ω is completely additive, i.e. ω( Ea ) = ω(Ea ) for any family {Ea } of mutually
a a
orthogonal projections in N,
11 Quantum Mechanics As a Theory of Observables and States (And. . . 263

(b) there is a density operator , i.e., a trace class operator on H with T r() = 1, such that
ω(A) = T r(A) for all A ∈ N.8

Condition (b) is often referred to as the “Born rule.”

11.3.2 From States to Probabilities, and from Probabilities

to States

A quantum state, whether pure or impure, normal or non-normal, induces a

probability measure on the projection lattice P(N): it is easily checked that if ω is
a state on N then Prω (E) := ω(E), E ∈ P(N), satisfies the conditions listed in §2
for a quantum probability measure. From Theorem 1, the probability Prω induced
by the state ω is completely additive iff ω is normal. Again, complete additivity
reduces to countable additivity when H is separable and to finite additivity when H
is finite dimensional.
Moving in the other direction, from probability measures on P(N) to states on
N, requires the use of Gleason’s theorem, which is the main technical result Itamar
cites in support of his T2. Below we will discuss the relevance of Gleason’s theorem
for T2, but in this section we confine ourselves to an exposition of the theorem. The
version of the theorem Itamar quotes applies only to ordinary non-relativistic QM
and, thus, in concert with the needed generalization of T1 the theorem needs to
be generalized to cover the branches of quantum theory that go beyond ordinary
QM. Thanks to twenty years of labor by a number of mathematicians the needed
generalization is available.
The original version of Gleason’s theorem for ordinary QM can be given the
following formulation:
Theorem 2 (Gleason). Let Pr be a quantum probability measure on P (B(H)) where H is
separable and dim(H) > 2. If Pr is countably additive then it extends uniquely to a normal
state on B(H).

This result has been generalized to cover non-separable Hilbert spaces and, more
significantly, quite general von Neumann algebras, so that it is applicable to all
branches of quantum theory:
Theorem 3 (Gleason generalized). Let N be a von Neumann algebra acting on a Hilbert
space H, separable or non-separable. Suppose that N does not contain any direct summands
of Type I2 . Then for any quantum probability measure Pr on P (N) there is a unique
extension of Pr to a quantum state ωPr on N. Further, the state ωPr is normal iff Pr is
completely additive.9

8 See Kadison and Ringrose (1987, Vol. 2, Theorem 7.1.12) and Bratelli and Robinson (1987,
Theorem 2.4.21).
9 This Theorem requires proof techniques that are different from the one used for the original

Gleason theorem; see Hamhalter (2003) and Maeda (1990) detailed treatments of this crucial
theorem. A Type I2 summand is isomorphic to the algebra of bounded operators on a two-
dimensional Hilbert space.
264 J. Earman and L. Ruetsche

11.3.3 The Born Rule and a Challenge for T2 and T2∗

The generalized Gleason’s theorem opens the door to those who want to endorse
T2 and relegate quantum states to bookkeeping devices across the board in quantum
theory. But why go through this door? One motivation comes from those who want
to apply Bruno de Finetti’s personalism to quantum theory and construe quantum
probabilities as the degrees of belief rational agents have about quantum events
P(N). Gleason’s theorem allows the personalists to claim that quantum states have
no independent ontological standing but function only to represent and track the
credences of Bayesian agents. This answer leads directly to Itamar’s T2: quantum
states are bookkeeping devices for quantum probabilities, which reflect rational
responsiveness to quantum event structures limned by T1.
Itamar writes: “The remarkable feature exposed by Gleason’s theorem is that the
event structure dictates the quantum probability rule. It is one of the strongest pieces
of evidence in support of the claim that the Hilbert space formalism is just a new
kind of probability theory” (Pitowsky 2003, p. 222). By “the quantum probability
rule” he means the Born rule. The Gleason theorem by itself does not yield the
Born rule for calculating probabilities and expectation values. The Born rule uses a
density operator representation of expectation values, and from the point of view of
states the validity of this representation comes from the restriction of possible states
to normal states; and from the point of view of probabilities the validity derives
from the restriction to completely additive measures.10 And it is exactly here that a
challenge arises for both of the competing theses T2 and T2∗.
The use of the Born Rule is, we claim, baked into the practice of applying
quantum theory. In the following section we offer evidence for this claim. The
challenge for the proponents of the opposing T2 and T2∗ is to explain and justify
this practice, or explain when and why the practice should be violated.

11.4 Justifying Normality/Complete Additivity

11.4.1 Normality and the Practice of QM

In this subsection we use ‘state’ in way that is neutral with respect to T2 vs T2∗;
so if you are a fan of T2 feel free to understand ‘state’ as qua representative of a

10 An awkwardness for Itamar’s account is the existence, in the case where dim(H) = 2, of
probability measures on P (B(H)) that don’t extend to quantum states on B(H). In order to
maintain that quantum states are just bookkeeping devices for the credences of rational agents
contemplating bets on quantum event structures, it seems that Itamar must either deny that two
dimensional Hilbert spaces afford quantum event structures, or articulate and motivate rationality
constraints (in addition to the constraint of conformity to the axioms of the probability calculus!)
that prevent agents from adopting “stateless” credences.
11 Quantum Mechanics As a Theory of Observables and States (And. . . 265

quantum probability measure. Our purpose is to indicate how the practice of QM

has baked into it the restriction to normal states.
Begin by noting that most standard texts on quantum theory use ‘state’ in a way
that makes it synonymous with normal state, e.g. it is assumed that expectation
values of observables follow the Born rule. A substantive example of the role of
normal states comes from the generally accepted analysis of quantum entanglement,
which is based on normal states. Let N1 and N2 be mutually commuting von
Neumann algebras acting on a common Hilbert space. Think of N1 and N2 as
characterizing the observables of two subsystems of a system whose algebra has
the tensor product structure N1 ⊗N2 . A state ω on the composite system algebra
N1 ⊗N2 is a product state iff ω(AB) = ω(BA) = ω(A)ω(B) for all A ∈ N1 and
B ∈ N2 . A normal state ω on N1 ⊗N2 is said to be quantum entangled over N1 and
N2 iff it cannot be approximated by convex linear combinations of normal product
states with respect to N1 and N2 , i.e. ω fails to lie in the norm closure of the hull
of convex linear combinations of normal product states. The following result links
entanglement and the non-abelian character of quantum observables:
Theorem 4 (Raggio 1988). Let N1 and N2 be mutually commuting von Neumann algebras
acting on a common Hilbert space. Then the following two conditions are equivalent:
(R1) At least one of N1 and N2 is abelian
(R2) No normal state on N1 ⊗N2 is entangled over N1 and N2 .

Conversely, the possibility of quantum entanglement requires that both subsystem

algebras have a quantum (= non-abelian) character. Thus, one can hold that the
key feature that separates classical from quantum physics is the difference between
abelian vs. non-abelian observables and at the same time agree with Schrödinger
that entanglement is the key feature separating quantum from the classical. This
result does not hold for non-normal states.
Such examples can be multiplied: check the theorems presented in textbooks
on quantum information and quantum computing and find that often the theorems
presuppose that ‘state’ means normal state. For future reference note how the
restriction to normal states affects the allowable pure states. When N=B(H), as
in ordinary QM sans superselection rules, the class of normal pure states is identical
with the class of vector states. In ordinary QM with superselection rules some vector
states are not pure—in particular, states that superpose across superselection sectors.
This, rather than the incorrect notion that superselection rules represent a limitation
on the superposition principle, is the way to bring home one of the implications of
superselection. When N is Type II or Type III there are no normal pure states; and
for Type III all normal states are vector states.
To say that the exclusive use of normal states is baked into the practice of
quantum theory is not to justify the practice, except insofar as ‘It works, so it must
be right’ is a justification. What one takes as a satisfying justification of the practice
depends upon whether one takes the side of Itamar’s T2 or our alternative T2∗. We
examine the contrasting justifications in turn.
266 J. Earman and L. Ruetsche

11.4.2 Justifying Normality: T2∗

Our T2∗ takes quantum states to codify objective, observer-independent features of

quantum system. From this perspective a justification of the exclusive reliance on
normal states would have to take the form of an argument showing that only normal
states are physically realizable. Here we give an example, drawn from the context
of the algebraic formulation of QFT, of such an argument.
AQFT postulates an association O → N(O) of von Neumann algebras N(O)
with open bounded regions O of Minkowski spacetime M satisfying the net
property, viz. if O ⊆ O then N(O) ⊆ N(O ). The quasi-local global algebra
N(M) is taken to be the smallest von Neumann algebra generated by the N(O) as
O ranges over all of the open bounded regions of M. Since Minkowski spacetime is
based on a Hausdorff manifold, for any spacetime point p ∈ M there is a sequence
{On }∞ ∞
n=1 of open regions such that On+1 ⊂ On for all n ∈ N and ∩n=1 On = {p}.
The property of local definiteness is designed to capture the notion that there are no
observables “at a point” so that physically realizable states become indistinguishable
in sufficiently small regions: technically, the requirement is that for any physically
realizable states ω and ϕ on N(M), ||(ω − ϕ)|N(On ) || → 0 as n → ∞ (see Haag
1992, p. 131). The only other assumption needed is the unobjectionable posit that
the physically realizable states on N(M) contain at least one normal state ϕ. It
follows that all physically realizable states ω are locally normal, i.e. for any p ∈ M
there is a neighborhood O  of p such that ω|  is normal (see Appendix 2 for a
N(O)
proof).
This result is modest since it gives no assurance that the neighborhood on which
the state is normal is large enough to accommodate an experiment one might hope
to conduct; and even such as it is its effectiveness turns on the plausibility of
the property of local definiteness. Nevertheless, it serves as a useful example of
how normality can be grounded in physical properties. Notice that, unless it can
be established a priori that there are no observables at a point, considerations of
rationality alone are inadequate to secure such a grounding.
There are much more ambitious arguments for normality that revolve around the
state preparation. We will postpone an examination of these arguments until §5.
But before reviewing these arguments we should also mention that on the other
side of the ledger there are occasional calls for the use of non-normal states. For
example, Srinvas (1980) argues that non-normal states are needed to define the joint
probabilities associated with successive measurements of quantum observables with
continuous spectra. Relatedly, Halvorson (2001) considers the use of non-normal
states as a means of assigning sharp values to the position of a particle and, thus, as
providing a basis for interpreting the modulus square |ψ(q)|2 of the wave function
(in the usual L2C realization) of a spinless particle as the probability density that the
particle has exact position q.11 Non-normal states also provide a means to overcome

11 For a contrary viewpoint, see Feintzeig and Weatherall (2019).

11 Quantum Mechanics As a Theory of Observables and States (And. . . 267

no-go results on the existence of conditional expectations for normal states such
as that of Takesaki (1972). Such considerations do not undermine our support
for T2∗; indeed they strengthen it by indicating that the sorts of considerations
physicists use to decide the admissibility of quantum states are quite different from
the considerations of rationality of credences the proponents of T2 would bring to
bear.

11.4.3 Justifying Normality: T2

From the point of view of T2 the justification for normality can only be that the
credence functions rational agents put on the projection lattice P(N) must be
completely additive and, thus, by the generalized Gleason theorem the states that
bookkeep these credences are normal.
One way to make good on the claim that the credence functions of rational agents
are completely additive is to mount a Dutch book argument. An agent who uses her
credence function to assign betting quotients to events is subject to Dutch book just
in case there is a family of bets, each of which she accepts but collectively have the
property that she loses money in every possible outcome. The goal is to show that
conforming credences to the axioms of probability with complete additivity is both
necessary and sufficient for being immune to Dutch book. The ‘only if’ part of the
goal can be attained if the agent is required to stand ready to accept any bet she
regards as merely fair, but not if she is only required to accept bets that she regards
as strictly favorable (see Skyrms 1992). This may be regarded as a minor glitch in
that for most applications of quantum theory a separable Hilbert space suffices; in
these cases complete additivity reduces to countable additivity, and the Dutch book
argument goes through even if the agent only accepts what she regards as strictly
favorable bets.
Nevertheless, personalists are on thin ice here since other constructions—such
as scoring rule arguments—personalists have used to try to prove that rational
credences should conform to the axioms of probability do not reach countable,
much less complete, additivity. Additionally, Bruno de Finetti, the patron saint
of the personalist interpretation of probability, adamantly denied that credences
need be countably additive much less completely additive (de Finetti 1972, 1974).
The only restrictions on admissible probability functions to which personalists
may appeal are restrictions justified by considerations of rationality alone. If the
patron saint of personalism is rational, it seems that the status of countable and
complete additivity is something about which rational parties can disagree. And if
countable/complete additivity can’t be rationally compelled then the restriction of
quantum probability measures to countably and completely additive ones—and ergo
268 J. Earman and L. Ruetsche

the restriction of quantum states to normal ones—eludes the reach of personalist

explanatory resources.12
Apart from the additivity issue the main challenge to the personalist interpreta-
tion of quantum probabilities comes from the implication that there are as many
quantum states as there are rational agents. But, we’ll suggest, the desideratum
that state preparation be explicable supports T2∗ against T2, and supports the
existence of a unique agent-independent state against the existence of as many
agent-dependent states as there are rationally admissible credence assignments.

11.4.4 Finitizing

We have been pressing the line that advocates of T2∗ have better resources for
justifying normality, as well as for explicating departures from normality, than
advocates of T2 have. But if normality requires no justification, this line evaporates,
taking with it our best grounds for preferring our own objectivist interpretation of
quantum probabilities to Itamar’s personalist interpretation of those probabilities.
And Itamar has at his disposal a maneuver that would eradicate the need to justify
normality. That maneuver is to finitize. In this subsection, we’ll sketch it briefly,
then explain why we think even Itamar should resist it.
Throughout his exposition of quantum mechanics as a theory of probability,
Itamar assumes the von Neumann algebras B(H)—the algebras framing quantum
bets and affording quantum event structures—to be finite dimensional. We can find
motivation for the assumption in his 2006: “A proposition describing a possible
event in a probability space is of a rather special kind. It is constrained by the
requirement that there should be a viable procedure to determine whether the event
occurs” (p. 217). This is because “only a recordable event can be the object of a
gamble” (p. 218). Itamar’s point is that the outcomes subject to quantum bets have
to be the sorts of thing we can settle. Otherwise we wouldn’t be able to resolve
the bets we’ve made. Given certain assumptions about our perceptual limitations
and our attention (and life) spans, this constrains quantum bets to concern events
with only finitely many possible outcomes—and thereby constrains quantum event
structures to be given by the projection lattices of finite dimensional von Neumann
algebras. And, of course, with quantum events structures thus constrained, complete
additivity collapses to finite additivity, and the rock-bottom requirement that
quantum probability functions be finitely additive implies that quantum states are
automatically normal. Normality requires no further justification; it makes no sense
to assess accounts of quantum probability on the basis of how well they justify
normality.

12 Bayesian statisticians are divided on the issue of additivity requirements. Some versions of
decision theory, such as that of Savage (1972), operate with mere finite additivity. Kadane et al.
(1986) argue that various anomalies and paradoxes of Bayesian statistics can be resolved by the
use of so-called improper priors and that such priors can be interpreted in terms of finitely additive
probabilities. For more considerations in favor of mere finite additivity see Seidenfeld (2001).
11 Quantum Mechanics As a Theory of Observables and States (And. . . 269

While we concede the formal merits of this maneuver, we contend that even
Itamar should resist it, because it fundamentally undermines the spirit and interest
of his interpretation of quantum mechanics. Baldly put, we think the finitizing
move is hubris. It amounts to demanding that the quantum event structure—the
topic of T1/T1∗, and the element all hands agree to be supplied by the physics—
conform to packages of gambles mortal human agents can entertain and settle.
That is, the finitizing maneuver sculpts the physical event structure by appeal to
the limitations of agents. We acknowledge and admire the resourcefulness and
integrity of personalist attempts to derive constraints on credences from features of
the agents entertaining them. But we take the present finitizing move to inject agents
where they don’t belong: into the physical event structures themselves. Itamar’s
conjunction of T1 and T2 is brilliant and provocative in no small part due to the
delicate division of labor it posits: physics supplies the event structure; we supply
everything else. Upsetting that division of labor, the finitizing move diffuses the
distinctive force of Itamar’s position. Without finitizing, anyone who agrees that
the quantum event structure is physical must take Itamar’s position seriously. With
finitizing, opponents might well complain that Itamar’s pulled a bait and switch, and
offered subjectified quantum event structures in lieu of physical ones.

11.5 Ontological Priority of States: State Preparation and

Objective Probabilities

To repeat, the view we are promoting is that quantum states codify objective features
of quantum systems, and the probabilities they induce are thereby objective. In
support of this view we offer the facts that (some) quantum states can be prepared
and that the probabilities they induce are borne out in the statistics of the outcomes
of experiments. Quantum theory itself contains the ingredients of a formal account
of state preparation.
The account contains two main ingredients, the first of which is the von
Neumann/Lüders projection postulate:
vN/L Postulate: If the pre-measurement state of a system with observable algebra N is ω and
a measurement of F ∈ P (N) returns a Yes answer, then the immediate post-measurement
ω(F AF )
state is ωF (A) := , A ∈ N, provided that ω(F ) = 0.13
ω(F )
Note that if ω is a normal state then so is ωF . The second ingredient is the concept
of a filter for states:
Def. A projection Fϕ ∈ P (N) is a filter for a normal state ϕ on N iff for any normal state ω
on N such that ω(Fϕ ) = 0,

13 von Neumann’s preferred version of the projection postulate differs from the version stated here
in the case of observables with degenerate spectra. Preliminary experimental evidence seems to
favor the Lüders version stated here; see Hegerfeldt and Mayato (2012) and Kumar et al. (2016).
270 J. Earman and L. Ruetsche

ω(Fϕ AFϕ )
ωFϕ (A) := = ϕ(A) for all A ∈ N.
ω(Fϕ )

In ordinary QM with N = B(H) the normal pure states are the vector states, and
as the reader can easily check any such state ψ has a filter Fψ ∈ P(B(H)), which
is the projection onto the ray spanned by the vector |ψ representing ψ. So suppose
that Fψ is measured and that the measurement returns a Yes answer. Then by the
vN/L postulate, if the pre-measurement state is ω then the post-measurement state
is ωFψ . And since Fψ is a filter for ψ the post-measurement state is ψ regardless of
what the pre-measurement state ω is, provided only that ω is a normal state and that
ω(Fψ ) = 0.
To repeat, this formal account of state preparation meshes with the statistics of
experiments. E.g. prepare a system in the normal pure state ψ. Immediately after
preparation make a measurement of some E ∈ P(B(H)) and record the result.
Repeat the preparation of ψ on the system or a similar system; then repeat the
measurement of E and record the result. Repeat . . . . Since the trials are independent
and identically distributed the strong law of large numbers implies that as the
number of trials is increased the frequency of Yes answers for E almost surely
approaches its theoretical value ψ(E). This is exactly what happens in actual
cases—at least there is always apparent convergence towards ψ(E).
The ansatz that only normal states are physically realizable has played a silent
but crucial role in the above account of state preparation. We have already suggested
that this ansatz fits more comfortably with our T2∗—which countenances physical
constraints on the space of admissible states—than it does with Itamar’s T2—
which countenances only rationality constraints. The restriction to normal states
has played a crucial rule because if a system were initially in a non-normal state
it would be inexplicable how it could be prepared in a normal state since neither
unitary evolution nor Lüders/von Neumann projection can cause a tiger to change
its stripes: if ω is non-normal then so is ω (•) = ω(U • U −1 ) for any unitary
ω(F • F )
U : H → H; and if ω is non-normal then so is ω (•) = for any
ω(F )
F ∈ P(B(H)) such that ω(F ) = 0 (see Ruetsche 2011). Turning this around, the
success in preparing systems in normal states provides a transcendental argument
for the ansatz. Transcendental arguments are a priori, if they are arguments at all.
So perhaps an unholy marriage of Kant and Itamar will avail advocates of T2 of the
resources necessary, on this story, to make sense of state preparation.
Another attitude toward the ansatz that only normal states are physically
realizable might be available to advocates of T2∗ embarrassed by transcendental
arguments but not embarrassed by certain dialectical moves made on behalf of
the Past Hypothesis (see e.g. Loewer 2012). This is the hypothesis that the initial
macrostate of the universe was exceedingly low entropy. Some fans of the Past
Hypothesis claim that (i) it helps explain prevalent time asymmetries integral to our
epistemic practices; that (ii) it thereby figures as an axiom in the simplest, strongest
systematization of physical fact; and that (iii) this qualifies the Past Hypothesis as a
physical law by the lights of the Best Systems Account of laws of nature. Despite
11 Quantum Mechanics As a Theory of Observables and States (And. . . 271

recoiling at this line of thought, we observe that less discriminating advocates of

T2∗ might adapt its basic architecture to argue for the lawlike status of a “Normality
Hypothesis,” that all physically realizable quantum states are normal. After all, the
Normality Hypothesis, by making sense of our central epistemic practice of state
preparation, earns its chops as an element of the simplest, strongest systematization
of physical fact, and so a physical law. And we also observe that this maneuver is not
available to advocates of Itamar’s T2. The maneuver explains the state restrictions
required to make sense of state preparation by appeal to physical law (the Normality
Hypothesis). Itamar would translate the restriction on physically realizable states
to a restriction on rationally admissible credence functions. But the explanation
of the restriction defies translation: physical laws and physical contingency can
explain physical restrictions, but only rationality constraints can explain restrictions
on credences.
In §7 below we will consider a difficulty in extending the account of state prepa-
ration developed above to more general von Neumann algebras of observables. But
in the next section we note the proponents of T2 and a personalist interpretation of
quantum probabilities can piggyback on the above formalism to give an alternative
account of state preparation. And in addition, they claim, this alternative account
dissolves the notorious measurement problem—a consideration that, if correct,
would give Itamar’s T2 a huge lead over our T2∗ in the race to make sense of
quantum theory’s undeniable empirical success.

11.6 Fighting Back: An Alternative Account of ‘State

Preparation’

11.6.1 Updating Quantum Probabilities

The proponent of T2 owes an alternative account of what happens in ‘state

preparation’ if states are just bookkeeping devices. Such an account starts with a
rule for updating probabilities on measurement outcomes. The generally accepted
updating rule for quantum probabilities uses Lüders conditionalization, which is
motivated by the following result:
Theorem 5 Let Pr be a completely additive quantum probability measure on the projection
lattice P (N) of a von Neumann algebra N which does not contain any direct summands
of Type I2 , and let F ∈ P (N) be such that Pr(F ) = 0. Then there is a unique functional
Pr(•//F ) on P (N) such that (a) Pr(•//F ) is a quantum probability, and (b) for all E ∈
P (N) such that E ≤ F , Pr(E//F ) = Pr(E)/ Pr(F ).14

Granted that conditions (a) and (b) capture features we want updating to have, the
uniquely determined updating rule specifies that if the pre-measurement probability

14 This is a slight generalization of the result of Cassinelli and Zanghi (1983).

272 J. Earman and L. Ruetsche

measure is Pr and if a measurement of F ∈ P(N) returns a Yes answer then the

updated post-measurement probability measure is Pr(•//F ).
To see what the updated measure comes to, note that the conditions of the
Theorem 5 make the generalized Gleason theorem applicable so that Pr extends
uniquely to a normal state ω on N. Then

ω(F EF ) ω(F EF )
Pr(E//F ) = = for all E, F ∈ P(N). (L)
ω(F ) Pr(F )

When E and F commute EF = F E = E ∧ F and (L) reduces to

Pr(EF ) Pr(E ∧ F )
Pr(E//F ) = = (B)
Pr(F ) Pr(F )

which is the rule of Bayes updating used in classical probability. However, when
E and F do not commute ω(F EF ) cannot be written as Pr(F EF ) since F EF ∈ /
P(N) and is not assigned a probability. It should be embarrassing to the proponent
of T2 that states have to be utilized in the updating rule, but let this pass.

11.6.2 ‘State Preparation’ According to T2

Call Xϕ ∈ P(N) a belief filter for a normal state ϕ on N iff for any completely
additive Pr on P(N) such that Pr(Xϕ ) = 0, Pr(E//Xϕ ) = ϕ(E) for all E ∈ P(N).
When the generalized Gleason theorem applies an application of (L) shows that
a state filter for ϕ is a belief filter for ϕ. The converse is also true: when the
generalized Gleason theorem applies a belief filter for ϕ is also a state filter for
ϕ (see Appendix 4 for a proof). The equivalence of belief filters and state filters
allows the proponents of T2 and a personalist interpretation of probability to claim
various virtues.
They can say that they can account for ‘state preparation’ without having to carry
the ontological baggage of T2∗. What the proponents T2∗ call a state filter for ϕ,
they call a belief filter for ϕ. And in virtue of Xϕ being a belief filter, all agents
who assign a non-zero prior probability to the belief filter Xϕ and who operate with
completely additive credence functions will, upon updating on the belief filter Xϕ ,
agree that posterior probabilities are equal to those assigned by ϕ. The state ϕ has
been ‘prepared’, but this just means that the bookkeeping device ϕ represents the
mutually shared updated opinion.
But please note that the personalist appropriation of §5’s formal account of state
preparation requires a restriction to completely additive probabilities—a restriction,
we’ve suggested, that fails to follow from considerations of rationality alone, unless
those considerations encompass transcendental arguments of a novel and surprising
nature. And, to review the discussion of §4.4, while the maneuver of finitizing
quantum event structures accomplishes the “restriction” in one fell swoop, it also
11 Quantum Mechanics As a Theory of Observables and States (And. . . 273

revises the division of labor so compellingly effected by Itamar’s T1 and T2. To our
minds, proponents of T2∗, who can invoke the dumb luck of physical contingency,
are better positioned to make sense of state preparation. Insofar as we’re assessing
proposals with respect to their capacity to deliver the goods of making sense of
quantum theory’s empirical success, this tips the balance in favor of our T2∗.
Upsetting the balance is the possibility that our strategy for making sense of
state preparation might not generalize beyond Type I von Neumann algebras. Given
our insistence on such generalization, such a possibility would hoist us on our own
petard. The next section gives life to the possibility, but also explains how we can
avoid being hoisted.

11.7 State Preparation in QFT

11.7.1 The Conundrum

The account of state preparation given in §5 focused on normal pure states. We have
already explained the restriction to normal states. The restriction to pure states is
explained by the fact that mixed states do not have filters and, thus, are outside the
ambit of the formal account of state preparation offered above (see Appendix 3 for a
proof). This fact leads a puzzle about how and whether state preparation is possible
in QFT.
It is widely assumed that local observers can only make measurements in
bounded regions of spacetime. If this is so it seems that such observers cannot
prepare global states, i.e. states on the quasi-local algebra N(M) for the entirely
of Minkowski spacetime M. If N(M) is Type I, as is typically the case, it admits
normal pure states, and such states have filters. But a state filter is a minimal
projection, and the local Type III algebras N(O) associated with open bounded
regions O ⊂ M do not contain any such projections. This is just as well, you may
say, for if a global state could be prepared by making measurements in a bounded
O ⊂ M some horrible form of non-locality would be in the offing. But things seem
just as bad at the local level. Since a local algebra N(O) is Type III it does not admit
any normal pure states, and filters for mixed states on N(O) do not exist in N(O).

11.7.2 The Buchholz, Doplicher, and Longo Theorem

A way out of this seeming conundrum was sketched by Buchholz et al. (1986) by
showing that, under certain conditions on the net O → N(O) of local algebras, local
state preparation is possible. The exposition begins with a definition:
Def. Let ϕ be a normal state on N(M). A projection Eϕ ∈ N(M) is a local filter for the
restriction ϕ|N(O) of ϕ to the local algebra N(O) iff for any normal state ω on N(M)

ω(Eϕ AEϕ ) = ω(Eϕ )ϕ(A) for all A ∈ N(O).

274 J. Earman and L. Ruetsche

We know that Eϕ cannot belong to N(O) itself. But ideally it can be located in some
N(O) where O  only slightly larger than O. So the hope is that for any normal state
ϕ on N(M) and any O ⊂ M there is a local filter Eϕ ∈ N(O)  for ϕ| 
N(O) with O
only slightly larger than O.
The hope for local filters can be fulfilled if the local algebras exhibit a property
called split inclusion:
Def. Let N1 and N2 be von Neumann algebras with N1 ⊆ N2 . The pair (N1 , N2 ) is a split
inclusion iff there is a Type I factor F such N1 ⊆ F ⊆ N2 .
 with
The net O → N(O) has the split inclusion property iff for any O and O
 
O ⊃ O, the pair (N(O), N(O)) is a split inclusion.
Theorem 6 (Buchholz et al. 1986). The split inclusion property of the local algebras is both
necessary and sufficient for the existence of local filters.

The proof of the sufficiency is given in Appendix 5.

Despite the importance of this theorem it is rarely mentioned in textbooks. An
exception is Araki’s Mathematical Theory of Quantum Fields. His summary of the
significance of the theorem:
The existence of the local filter for a state in a certain domain means that the state can be
prepared by using some apparatus which is a little larger than the original one, and this
assumption is equivalent with the assumption of split inclusion. (Araki 2009, 191).

In terms of our discussion the moral is that the competing stories the proponents of
T2 and T2∗ told in the preceding section about state preparation in ordinary QM can
be retold in QFT for states on local algebras just in case the split inclusion property
holds.15
So what is the status of the split inclusion property? It is known to hold in some
models of AQFT and fail in others. That the dividing line is being crossed is signaled
by the number of degrees of freedom in the model. It is often assumed that the
vacuum state for N(M) is a vector state whose generating vector is cyclic and
separating. Halvorson (2007) shows
Theorem 7. Let O → N(O) be a net of von Neumann algebras acting on H, and suppose
that there is a vector |Ω ∈ H that is cyclic and separating for all the local algebras in the
net. If the split inclusion property holds then H is separable.16

Physical models of AQFT feature vacuum states that are cyclic and separating.
Further, the split inclusion property is known to be entailed by the nuclearity
condition, which expresses a precise form of the idea that as the energy of a system
increases the energy level density should not increase too rapidly (see Buchholz and
Wichmann 1986). Perhaps a case can be made that the split inclusion property picks
out “physically reasonable” models of AQFT, but it is clearly a contingent property.

15 One of us (LR) has reservations about the interpretational morals being drawn from the BDL
theorem; see Appendix 6.
16 Let N be a von Neumann algebra acting on H. A vector |Ω ∈ H is cyclic iff N|Ω is dense in

H. It is separating vector iff A|Ω = 0 implies A = 0 for any A ∈ N such that [A, N] = 0.
11 Quantum Mechanics As a Theory of Observables and States (And. . . 275

Of course, its contingency isn’t a challenge for Itamar’s T2. The property restricts
which algebras afford QFT’s quantum event structure. The topic of Itamar’s T1 and
our friendly amendment thereof to T1∗, quantum event structures are something
Itamar regards as a matter of physics.

11.8 Taking Stock

The proponents of T2 might claim to have achieved a standoff with the proponents
of T2∗. In ordinary QM where filters exist for all normal pure states the argument
for T2∗ from state preparation can be rewired to give an alternative account of ‘state
preparation’ in terms of belief fixation that is in accord with T2. In QFT the situation
is a bit more complex. At the global level filters for states on the global algebra are
unavailable to local observers so that the state preparation argument for T2∗ falters.
At the local level the story divides. In models where the split inclusion property
holds there are local filters and, thus, there is a standoff over local state preparation
parallel to the situation in ordinary QM; and in models where the split inclusion
property fails there are no local filters so that the state preparation argument for T2∗
falters.
The proponents of T2∗ may wish to put a different spin on matters. At the global
level in AQFT there is no preparation story for states on N(M) for the personalist
to piggyback on. The proponents of T2∗ can not only admit this but explain why
measurements by local observers cannot be used to prepare global states. And they
will also note that insofar as the practice of physics presupposes that there is a unique
global state, the proponents of T2 are at a loss to explain the practice. Similarly, at
the local level there is no preparation story for local states when the split inclusion
property fails. Again proponents of T2∗ can explain why measurements by local
observers cannot be used to prepare local states; and they will also note that insofar
as the practice of physics presupposes that there is on any local N(O) a unique local
state, the proponents of T2 are at a loss to explain the practice.
Setting aside QFT impediments, we take the proponents of T2∗ to have an
advantage over the proponent of T2. Both the formal T2∗ and the appropriated T2
accounts of state preparation by filtration need to restrict admissible quantum prob-
ability functions to completely additive ones. The expanded explanatory resources
afforded proponents of T2∗ give them the means to do so. Proponents of T2 need
to anchor such restrictions in considerations of rationality alone. It is not at all clear
that such considerations suffice.
Of course, if proponents of T2 could solve the measurement problem, there
would be good reason to set such quibbles aside. The next section contends that
they cannot.
276 J. Earman and L. Ruetsche

11.9 The Measurement Problem

The von Neumann/Lüders projection postulate gives the empirically correct pre-
scription for calculating post-measurement probabilities. If, following T2∗, quan-
tum states codify objective features of quantum systems and if, following John
Bell’s injunction, ‘measurement’ is not to be taken as primitive term of quantum
theory then there is a problem in accounting for the success of the vN/L postulate.
For numerous no-go results show that treating a quantum measurement as an
interaction between an apparatus and an object system and describing the composite
apparatus-object system interaction as Schrödinger/Heisenberg evolution does not
yield the desired result.
It is tempting to think that the problem can be dissolved by rejecting Bell’s
injunction and also rejecting T2∗ in favor of T2. For these twin rejections open
the way to say: The change of state postulated in vN/L does take place but this
change is innocuous and calls for no further explanation; for the change is simply
a change in the bookkeeping entries used to track the change in credence functions
when updating on measurement outcomes. In more detail, when the information
that a measurement of F ∈ P(N) gives a Yes result is acquired a completely
additive credence function Pr on P(N) is updated by Lüders conditionalization on
F to Pr (•) = Pr(•//F ). Assuming that the generalized Gleason theorem applies,
it follows that there are unique normal states ω and ω bookkeeping Pr and Pr
respectively, and it follows from the definition of Lüders conditionalization that
ω(F • F )
these bookkeepers are related by ω (•) = , just as vN/L requires. That’s
ω(F )
all there is to it.17
We demur: there is more, and the more steers us back towards the original
measurement problem. On T2∗ there is a unique but, perhaps, unknown pre-
measurement state, and the vN/L postulate refers to the change in this state. By
contrast, on T2 there are as many different bookkeeping states as there are agents
with different pre-measurement credence functions, and unless the F in question
is a belief filter there are many different post-measurement bookkeeping states.
Only one of these post-measurement bookkeeping states gives correct predictions
about observed frequencies, something easily understood on T2∗ since only one of
these states induces the objective chances of post-measurement results. Thus, on
T2 there is the 0th measurement problem of how to explain uniqueness of states:
how, in the parlance of T2, to explain why exactly one of the myriad admissible
probability functions available accords with subsequent observation. The problem
becomes particularly pressing in cases where there are no belief filters, as in models
of QFT where the split property fails.

17 Or rather that is all there is supposed to be to what Pitowsky (2003) and Bub and Pitowsky (2010)

call the ‘big’ measurement problem. They think that after the big problem is dissolved there still
remains the ‘small’ measurement problem, viz. to explain the emergence of the classical world we
observe.
11 Quantum Mechanics As a Theory of Observables and States (And. . . 277

The 0th measurement problem threatens to spill over into the BIG measurement
problem. It is our experience that after measurements, the credences available to
observers narrow to those corresponding to full belief in one outcome or another.
That is, upon updating on a measurement outcome, I obtain a credence assigning
probability 1 to an element of the set {Ei } of spectral projections of the pointer
observable. I never obtain a credence assigning probability 1 to a member of a set of
projections {Ei } incompatible with the pointer observable. It is not clear how Itamar
can explain why our options are invariably narrowed in these characteristic ways.
And even if he can, he faces a further mystery: why does a measurement deliver me
to full belief in one rather than another of the {Ei }? “Because that outcome—En
(say) was observed!” he might reply, rehearsing §6’s account of belief fixation via
filtration. But §6’s account posits an outcome on which all believers update—and it
is not clear that Itamar has the wherewithal to deliver the posit. Concerning quantum
event structures, T1 says nothing about which quantum events are realized in any
given physical situation. Thus it’s unclear that Itamar can say what the physical facts
are, much less that they include measurement outcomes. Compare our view, which
can marshal the physical quantum state to characterize the facts. The physical state,
along with some account of what’s true of a system in that state (e.g. the unjustly
maligned eigenvector-eigenvalue link) is the right kind of resource for saying what
the physical facts are. Alas, to date, no one’s articulated a rule that manages, without
legerdemain, to number measurement outcomes among the physical facts.
Additionally, even if ‘measurement’ is taken as a primitive it makes sense to ask
how physically embodied agents acquire information about measurement results.
Are we to take this as an unanalyzed process? Or do we search for an account using
quantum theory to analyze the interaction of agents with the rest of the world? The
former strengthens the suspicion that real problems are being swept under the rug.
The latter points to rerun of the original problem.

11.10 Conclusion

The key question separating our position from Itamar’s is how to understand
quantum probabilities and quantum states. Are they a matter of physics (our T2∗)
or a matter of rationality (Itamar’s T2)? But there is another question in the vicinity,
illuminating further points of contention. The question is: just how much realism
about QM is possible?
One way to approach the question is see how well candidate realisms accommo-
date the key explanandum that QM works. A stability requirement for a would-be
abductive realist is that what she believes when she believes QM must be consistent
with the evidence she cites to support her belief in QM. That evidence takes the
form of a history of preparing physical systems, attributing them quantum states,
extracting probabilistic predictions from those states, and verifying those predic-
tions by means of subsequent measurements. Interpretations that render preparation
infeasible, statistical prediction unreachable, or measurement impossible fail this
278 J. Earman and L. Ruetsche

stability requirement. They are self-undermining. If an interpretation of QM is self-

undermining, belief in QM under that interpretation is more realism about QM than
is possible. (Although we concede that the foregoing applies only to abductively-
supported realisms, we consider realisms motivated by other means to be tasteless.)
Itamar takes his subjective interpretation to afford as much realism as you can
have about QM. He calls it “realism about quantum gambles.” We can explicate it
as realism about QM understood in terms of T1 and T2: quantum event structures
are physically real; quantum probabilities are all in our heads. One could be even
less of a realist about QM than Itamar is. One might, with Bohmians, deny T1 by
asserting physical events to have some structure different from that given by the
lattice of closed subspaces of Hilbert space.18 But you can’t be more of a realist,
Itamar cautions. One way to be more of a realist is to believe QM understood in
terms of T1 (and its generalization T1∗) and T2∗. On this doubly realist view, both
quantum event structures and quantum probabilities/states are physically real. For
Itamar, the reason you can’t have this much realism is the measurement problem.
Understanding quantum states as physical renders understanding measurement
impossible. A realism guided by T1∗ and T2∗ is self-undermining.
We have suggested here that both our line and Itamar’s fail stability, but for
interestingly different reasons. Itamar’s fails to accommodate state preparation, and
fails because rationality considerations are the only grounds a personalist can give
for restricting admissible credences. Restricting admissible states to normal states
(and bracketing momentarily complications encountered in QFT), we can invoke
vN/L to make sense of state preparation (§5). Taking states to be physical, we have
the resources to justify the restriction we need: the restriction might obtain as a
matter of physical contingency or as matter of metaphysically contingent physical
law. Itamar’s personalist can co-opt the account of state preparation developed in §5.
But the co-option requires restricting admissible credences to completely additive
ones. And this restriction is heterodox when gauged against the personalist gospel
according to de Finetti. This leaves it contentious at best whether it’s a constraint
dictated by the demands of rationality alone.
We confess that QFT poses impediments to the account of state preparation
available to advocates of T2∗. §7 details how those advocates might deal with them.
Advocates of T2 have an additional maneuver by which they might negotiate these
same impediments. That maneuver is the one discussed in §4.4., of coarse-graining
and/or finitizing in a way that bleaches physical significance from the infinitary and
otherwise exotic algebras encountered in the setting of QFT. The finitizing move
also helps T2 advocates cope with preparation in more homely settings. Collapsing
countable to finite additivity, it effects the restriction on credences needed, only by
something like fiat. The fiat is the price here, and we don’t buy it.

18 Thisis hardly to deny that Bohmians are realists. They’re just realists about something else—the
Bohm theory. Standard QM stands to the Bohm theory as an effective theory stands to a more
fundamental theory whose coarse-grained implications it mimics. Thus realism about the Bohm
theory induces something like “effective realism” about QM. The point is that this represents less
realism about the theory than Itamar’s position. For more on effective realism, see Williams (2017).
11 Quantum Mechanics As a Theory of Observables and States (And. . . 279

What about the measurement problem? Both views encounter it in some form.
Both are self-undermining. But at least our view does not undermine physics. For
advocates of T2∗, the measurement problem signals that something physical is
eluding our grasp—and that signal is a call to develop new physics. Advocates
of T2 deny that anything is missing from the physics. Indeed, they contend that
the measurement problem results from putting something in the physics—quantum
states that reflect objective probabilities—that doesn’t belong there. Taking the
measurement problem to be thereby dissolved, they predict that no fruitful new
physics is to be had by confronting the problem.
In the end, one’s attitude towards the measurement problem comes down to a bet
on what is the most fruitful way to advance physics. Our bet is that path starts from
taking the measurement problem as a problem to be solved rather than dissolved;
and further, our bet is that advancement down this path will require new physics,
perhaps something along the lines of a stochastic mechanism to produce the changes
postulated by vN/L or perhaps a radically new theory.

Appendix 1 The Lattice of Projections

A projection E ∈ N is an “observable,” i.e. a self-adjoint operator, and it is

idempotent, i.e., E 2 = E. The collection of projections P(N) is equipped with
a natural partial order, viz. for E, F ∈ P(N), F ≤ F iff range(E) ⊆ range(F ).
P(N) is a lattice because it is closed under meet ∧ (the least upper bound) and
join ∨ (greatest lower bound) in this partial order. E1 ,E2 ∈ P(N) are mutually
orthogonal iff E1 E2 = E2 E1 = 0 (the null projection). When E1 and E2
are mutually orthogonal E1 ∨ E2 = E1 + E2 . Complementation in P(N) is
orthocomplementation, i.e. E c = E ⊥ = I − E.

Appendix 2 Locally Normal States in Algebraic QFT

Assume that among the physically realizable states on the quasi-local global algebra
N(M) there is at least one normal state ϕ. Assume also the property of local
definiteness. It follows that any physically realizable state ω is locally normal, i.e.
for any p ∈ M there is a neighborhood O  of p such that ω|  is normal. To
N(O)
∞
see this, consider a sequence {On }n=1 of open neighborhoods in M descending
to p. By local definiteness there must be a sufficiently large n0 ∈ N such that
||(ω − ϕ)|N(On ) || < 2. This implies that ω|N(On ) and ϕ|N(On ) belong to the
0 0 0
same folium, and since ϕ|N(On ) is normal so is ω|N(On ) . To complete the proof
0 0
 = On0 .
set O
280 J. Earman and L. Ruetsche

Appendix 3 No Filters for Mixed States

As a preliminary note that if ω is a state on N such that ω(E) = 1 for E ∈ P(N)

then it follows from the Cauchy-Schwartz inequality that ω(EF E) = ω(F ) for all
F ∈ P(N).
Now let Fϕ ∈ P(N) be a filter for a normal state ϕ on N. Since Fϕ = O
there must be a normal state ω such that ω(Fϕ ) > 0. Then from the filter property
ω(Fϕ Fϕ Fϕ )
= 1 = ϕ(Fϕ ). Suppose that ϕ is a mixed that: ϕ := λ1 φ 1 + λ2 φ 2 ,
ω(Fϕ )
where 0 < λ1 , λ2 < 1 and λ1 + λ2 = 1. Since ϕ(Fϕ ) = 1 it must be the case
that φ 1 (Fϕ ) = φ 2 (Fϕ ) = 1. Now consider some other normal mixed state ϕ  :=
λ1 φ 1 + λ2 φ 2 with λ1 = λ1 and λ2 = λ2 . Note that ϕ  (Fϕ ) = 1 and, therefore (as
seen above), ϕ  (Fϕ EFϕ ) = ϕ  (E) for all E ∈ P(N). Using again the filter property,
ϕ  (Fϕ EFϕ )
= ϕ  (E) = ϕ(E) for all E ∈ P(N). Since N is the weak closure
ϕ  (Fϕ )
of P(N) and since normal states are weakly continuous it follows that φ  = φ, a
contradiction.

Appendix 4 Belief Filters Are State Filters

Suppose that Xϕ ∈ P(N) a belief filter for the normal state ϕ, and let Pr be a
completely additive probability on P(N) such that Pr(Xϕ ) = 0. By the generalized
Gleason theorem it follows that ω(Xϕ ) = 0 and that
ω(Xϕ EXϕ )
= ϕ(E) for all E ∈ P(N) (*)
ω(Xϕ )
where ω is the normal state that extends Pr. Now use the facts that N is the weak
closure of P(N) and that normal states are continuous in the weak topology to
conclude that
ω(Xϕ AXϕ )
= ϕ(A) for all A ∈ N (**)
ω(Xϕ )
Since Pr was arbitrary and since completely additive probabilities are in one-
one correspondence with normal states, (**) holds for any normal state ω such that
ω(Xϕ ) = 0, which is to say that Xϕ is a state filter for ϕ.

Appendix 5 The Buchholz-Doplicher-Longo Theorem

Here we sketch the proof that the split inclusion property is sufficient for the
existence of local filters in AQFT; for a proof of the converse the reader is referred
to Buchholz et al. (1986),
11 Quantum Mechanics As a Theory of Observables and States (And. . . 281

Consider any open O ⊂ M. By the split inclusion property of the net O →

 is even slightly larger than O there is a Type I
N(O) of von Neumann algebras, if O

factor F such that N(O) ⊆ F ⊆ N(O). If E ∈ F is a minimal projection for F then

EAE = C(A)E for all A ∈ F (11.1)

where C(A) is a complex number depending on A. The support projection Eψ for a

normal pure (= vector) state ψ on F B(H) is the projection onto the ray spanned
by the vector |ψ ∈ H representing ψ. Since Eψ is a minimal projection and since
ψ(Eψ ) = 1 and ψ(Eψ AEψ ) = ψ(A), it follows from (11.1) that ψ(Eψ AEψ ) =
ψ(A) = C(A) and, thus,

Eψ AEψ = ψ(A)Eψ for all A ∈ F. (11.2)

We can conclude that for any normal state ω on N(M)

ω(Eψ AEψ ) = ω(Eψ )ψ(A) for all A ∈ F. (11.3)

And a fortiori (11.3) holds for all A ∈ N(O) ⊂ F. The upshot is that if ϕ is a
normal state on N(M) whose restriction ϕ|N(O) to the local algebra N(O) extends
to a normal pure state ψ on F then the support projection Eψ for ψ serves as a local
filter for ϕ and, as hoped, Eψ belongs to N(O)  associated with a O  only slightly
larger than O. It remains only to note that any normal state on N(O) extends to a
vector state on F. Since N(O) is Type III, any normal state ξ on N(O) is a vector
state. So there is a |ξ  ∈ H such that ξ (A) = ξ |A|ξ  for all A ∈ N(O). This state
obviously extends to the vector state ξ̄ on F B(H) where ξ̄ (B) = ξ |B|ξ  for all
B ∈ F. The support projection for this state Eξ̄ ∈ F ⊂ N(O)  is then the desired
local filter.

Appendix 6 Interpreting the Buchholz-Doplicher-Longo

Theorem

While the Buchholz-Doplicher-Longo theorem is airtight, its physical interpretation

poses some puzzles. No filter for any state on N(O) lies inside that algebra, which
is assumed to contain all observables pertinent to the region O. Even in the best
case that a filter for a state on N(O) belongs to an algebra associated with a slightly
larger region that contains O, this has the odd consequence that an experimenter
who wants to prepare her laboratory, assumed to be coextensive with O, in a state
has to leave her laboratory in order to succeed. This predicament might strike some
as in tension with the spirit of local quantum physics. A further oddity is the fact
that, while Eϕ ∈ N(O)  might be a filter for normal state ϕ on N(O), it is not a filter
for any normal state on its home algebra N(O).  A Type III algebra, N(O) admits no
282 J. Earman and L. Ruetsche

filters for its normal states. And this short-circuits what might otherwise have been
a comforting physical story about Buchholtz-Doplicher-Longo preparations. The
comforting story is that we are preparing a state ϕ on N(O) by preparing a state on
a larger algebra containing N(O); ϕ is just the restriction to N(O) of the prepared
superstate. (The comforting story incidentally explains why, contrary to ordinary
expectations, the states we prepare in QFT are mixed. The explanation is that in the
presence of entanglement, subsystem states are inevitably mixed, and subsystem
states are what BDL preparations prepare.) The comforting story is short-circuited
because in the BDL scenario, we aren’t preparing a state on the larger algebra N(O)
 
/the larger region O. We aren’t because we can’t: normal states on N(O) lack filters
 While such puzzles are no reason to disbelieve the theorem, they leave some
in O.
of us wondering, sometimes, what we’re believing when we believe this piece of
mathematics.

Acknowledgements This work reflects a collaboration with the late Aristides Arageorgis extend-
ing over many years. We are indebted to him. And we thank Gordon Belot, Jeff Bub, and the editors
of this volume for valuable guidance and feedback on earlier drafts.

References

Araki, H. (2009). Mathematical theory of quantum fields. Oxford: Oxford University Press.
Bratelli, O., & Robinson, D. W. (1987). Operator algebras and quantum statistical mechanics I
(2nd ed.). Berlin: Springer.
Bub, J., & Pitowsky, I. (2010). Two dogmas of quantum mechanics. In S. Saunders, J. Barrett, A.
Kent, & D. Wallace (Eds.), Many worlds? Everett, quantum theory, & reality (pp. 433–459).
Oxford: Oxford University Press.
Buchholz, D., & Wichmann, E. H. (1986). Causal independence and the energy-level density of
states in local quantum field theory. Communications in Mathematical Physics, 106, 321–344.
Buchholz, D., Doplicher, S., & Longo, R. (1986). On Noether’s theorem in quantum field theory.
Annals of Physics, 170, 1–17.
Cassinelli, G., & Zanghi, N. (1983). Conditional probabilities in quantum mechanics. Nuovo
Cimento B, 73, 237–245.
de Finetti, B. (1972). Probability, induction and statistics. London: Wiley.
de Finetti, B. (1974). Theory of probability (2 Vols). New York: Wiley.
Drish, T. (1979). Generalizations of Gleason’s theorem. International Journal of Theoretical
Physics, 18, 239–243.
Eilers, M., & Horst, E. (1975). Theorem of Gleason for nonseparable Hilbert spaces. International
Journal of Theoretical Physics, 13, 419–424.
Feintzeig, B., & Weatherall, J. (2019). Why be regular? Part II. Studies in History and Philosophy
of Science Part B: Studies in History and Philosophy of Modern Physics, 65, 133–144.
Haag, R. (1992). Local quantum physics. Berlin: Springer.
Halvorson, H. (2001). On the nature of continuous physical quantities in classical and quantum
mechanics. Journal of Philosophical Logic, 30, 27–50.
Halvorson, H. (2007). Algebraic quantum field theory. In J. Earman & J. Butterfield (Eds.), Hand-
book of the philosophy of science, philosophy of physics part A (pp. 731–864). Amsterdam:
Elsevier.
Hamhalter, J. (2003). Quantum measure theory. Dordrecht: Kluwer Academic.
11 Quantum Mechanics As a Theory of Observables and States (And. . . 283

Hegerfeldt, G. C., & Mayato, R. (2012). Discriminating between von Neumann and Lüders
reduction rule. Physical Review A, 85, 032116.
Kadane, J. B, Schervish, M. J., & Seidenfeld, T. (1986). Statistical implications of finitely additive
probability. In P. K. Goel & A. Zellner (Eds.), Bayesian inference and decision techniques (pp.
59–76). Amsterdam: Elsevier.
Kadison, R. V., & Ringrose, J. R. (1987). Fundamentals of the theory of operator algebras (2 Vols).
Providence: Mathematical Society.
Kumar, C. S., Shukla, A., & Mahesh, T. S. (2016). Discriminating between Lüders and von
Neumann measuring devices: An NMR investigation. Physics Letters A, 380, 3612–2616.
Loewer, B. (2012). Two accounts of laws and time. Philosophical Studies, 160, 115–137.
Maeda, S. (1990). Probability measures on projections in von Neumann algebras. Reviews in
Mathematical Physics, 1, 235–290.
Petz, D., & Rédei, M. (1996). John von Neumann and the theory of operator algebras. The
Neumann Compendium, 163–185.
Pitowsky, I. (2003). Betting on the outcomes of measurements: A Bayesian theory of quantum
probability. Studies in the History and Philosophy of Modern Physics, 34, 395–414.
Pitowsky, I. (2006). Quantum mechanics as a theory of probability. In: W. Demopoulos & I.
Pitowsky (Eds.), Physical theory and its implications; essays in honor of Jeffrey Bub (pp. 213–
240). Berlin: Springer.
Raggio, G. A. (1988). A remark on Bell’s inequality and decomposable normal states. Letters in
Mathematical Physics, 15, 27–29.
Ruetsche, L. (2011). Why be normal? Studies in History and Philosophy of Modern Physics, 42,
107–115.
Savage, L. J. (1972). Foundations of statistics (2nd ed.). New York: Wiley.
Seidenfeld, T. (2001). Remarks on the theory of conditional probability. In V. F. Hendricks, et al.
(Eds.), Probability theory (pp. 167–178). Dordrecht: Kluwer Academic.
Skyrms, B. (1992). Coherence, probability and induction. Philosophical Issues, 2, 215–226.
Srinvas, M. D. (1980). Collapse postulate for observables with continuous spectra. Communica-
tions in Mathematical Physics, 71, 131–158.
Takesaki, M. (1972). Conditional expectations in von Neumann algebras. Journal of Functional
Analysis, 9, 306–321.
Williams, P. (2017). Scientific realism made effective. The British Journal for the Philosophy of
Science, 70, 209–237.
Chapter 12
The Measurement Problem and Two
Dogmas About Quantum Mechanics

Laura Felline

Abstract According to a nowadays widely discussed analysis by Itamar Pitowsky,

the theoretical problems of QT are originated from two ‘dogmas’: the first,
forbidding the use of the notion of measurement in the fundamental axioms
of the theory; the second, imposing an interpretation of the quantum state as
representing a system’s objectively possessed properties and evolution. In this paper
I argue that, contrarily to Pitowsky analysis, depriving the quantum state of its
ontological commitment is not sufficient to solve the conceptual issues that affect
the foundations of QT.
In order to test Pitowsky’s analysis I make use of an argument elaborated by Amit
Hagar and Meir Hemmo, showing how some probabilistic interpretations of QT fail
at dictating coherent predictions in Wigner’s Friend situations. More specifically, I
evaluate three different probabilistic approaches: qBism, as a representative of the
epistemic subjective interpretation of the quantum state; Jeff Bub’s information-
theoretic interpretation of QT, as an example of the ontic approach to the quantum
state; Itamar Pitowsky’s probabilistic interpretation, as an epistemic but objective
interpretation. I argue that qBism maintains self-consistency in Wigner’s Friend
scenarios, although, in the resulting picture, the real subject matter of QT clashes
alarmingly with scientific practice. The other two approaches, instead, strictly fail
when confronted with Wigner’s Friend scenarios.

Keywords Measurement · Quantum state · Wigner’s Friend · Quantum

information theory · Quantum Bayesianism

L. Felline ()
Department of Philosophy, University of Roma Tre, Rome, Italy

© Springer Nature Switzerland AG 2020 285

M. Hemmo, O. Shenker (eds.), Quantum, Probability, Logic,
Jerusalem Studies in Philosophy and History of Science,
https://doi.org/10.1007/978-3-030-34316-3_12
286 L. Felline

12.1 Introduction

There is a time-honoured approach to the foundations of Quantum Theory (QT),

according to which the oddities of this theory originate from the improper inter-
pretation of its subject study as concerning properties and behaviour of material
objects, being them corpuscles or waves, represented by the quantum state. This
interpretation is often held responsible, for instance, for the difficulties in explaining
quantum correlations, or the measurement problem.
The argument behind this association is quite intuitive. Let’s take the problem
of (apparently) non-local quantum correlations, and the illustrative case of an
EPR-Bohm entangled pair of particles A and B, which are sent apart. The EPR-
Bohm state does not associate a determinate property to either particle, however,
when a measurement on A is performed, the entangled state of A + B effectively
and instantaneously collapses, associating now a specific value to the measured
observable, not only in the system A, but in the system B as well. As a result, due
to A’s and B’s perfect anticorrelation, knowing the result of a measurement on A
automatically also informs us about the state of B.
Now, if the quantum state is interpreted realistically, as encoding the properties
possessed by a system, and its evolution, then a straightforward interpretation takes
the collapse as a physical change both in A’s and in B’s properties. But B was
never disturbed by the observer’s interaction with A, so an explanation is in order
for B’s behaviour. At the origin of the difficulty of the sought-after explanation is
the tension between the natural assumption that the measurement over A has in
some way caused the change in B’s state, and the fact that causal processes can’t be
instantaneous.
This difficulty seems to disappear once we give up the assumption that quantum
states, contrarily to classical states, represent properties and evolution of physical
systems. If the quantum state is just an economical codification of a probability
distribution over measurement results, then the collapse of the wave function does
not represent an actual physical process, and B’s state change does not have to be
explained physically.
So, according to this argument, the problem of non-locality (and, as we are about
to see, the measurement problem with it) is explained away once one rejects what I
will call from now on a representational function of the quantum state.1
But this conclusion would be too quick. For instance, in the case of non-locality,
a genuine solution to the problem must necessarily provide an explanation of why,
given that there is no physical process connecting the occurrence of A’s and B’s
measurement results, said results are always perfectly anti-correlated. But this is not
an easy task. As a matter of fact, when it comes to provide an alternative explanation
to quantum conundrums, non-representational approaches plod along as well.

1 See later in this section ad especially note 2 for some clarifications about the terminology.
12 The Measurement Problem and Two Dogmas About Quantum Mechanics 287

Because of these persisting difficulties, mainstream physics settles for the

adoption of a pragmatic but inconsistent attitude towards the quantum state (see
Wallace 2016), to be interpreted as representing a system or as a codification of
probabilities of measurement results, depending on what works in that situation.
This, for instance, is physicist David Mermin, about the two-slit experiment:
We know what goes through the two slits: the wave-function goes through, and then
subsequently an electron condenses out of its nebulosity onto the screen behind. Nothing
to it. Is the wave-function itself something real? Of course, because it’s sensitive to the
presence of both slits. Well then, what about radioactive decay: does every unstable nucleus
have a bit of wave-function slowly oozing out of it? No, of course not, we’re not supposed
to take the wave-function that literally; it just encapsulates what we know about the alpha
particle. So then it’s what we know about the electron that goes through the two slits?
Enough of this idle talk: back to serious things! (Mermin 1990, p. 187, cited in Maudlin
1997, p. 146)

This being said, in the last decades non-representational approaches to QT have

experienced new life thanks to the evolutions of Information Theory. The hope
is that the latter might provide a formal background powerful enough to allow a
genuine solution to the theoretical problems of QT, without necessarily collapsing
into instrumentalism. Accordingly, QT is not about particles or waves and their
behaviour, but is rather about information.
A terminological clarification is in order at this point. The approaches denying
that QT is about properties and behaviour of physical entities are often labelled
anti-realist approaches to QT (see e.g. Myrvold 2018, §4.2); however, this charac-
terization might be confusing, since not every such account is anti-realist. In fact,
the majority of them openly rejects anti-realism (and this is definitely true of the
three approaches examined in this paper): QT is not a theory about the elementary
constituents of matter, but it is still realist with respect to its subject matter.
In this paper, I adopt Jeff Bub’s terminology and label the two competing views
probabilistic and representational approaches, where the former interprets QT as
a theory of probability (and where we can further distinguish between an ontic
and an epistemic interpretation of the quantum state), while the latter interpret QT
as a theory about the elementary constituents of matter, their properties and their
dynamical behaviour.2
As a more general point, I argue that what was already shown for more traditional
probabilistic approaches, equally applies to the new information-based formula-
tions: contrarily to the simplistic attitude often adopted even in recent illustrations
of the subject, depriving the quantum state of its ontological commitment is not
sufficient to solve the conceptual issues that affect the foundations of QT.

2A further clarification: Bub takes this terminology from Wallace (2016) where, however,
probabilistic interpretations are characterized as those according to which observables always have
determinate values, while the quantum state is merely “an economical way of coding a probability
distribution over those observables” (p. 20). I think this characterization is incorrect of the three
approaches here analysed, since they do not take observables as determinate all the time, therefore
I don’t refer to Wallace terminology.
288 L. Felline

In arguing for this general point, I focus on the measurement problem, typically
taken to be explained away in non-representational accounts. As in the above
illustrated example of non-locality, and contrarily to what often claimed by the
literature, also in the case of the measurement problem, the majority of non-
representational accounts fail to provide a genuine solution.
One of the most systematic expositions of the non-representational view and of
its relationship with the measurement problem was put forward few years ago by
Itamar Pitowsky, in a series of paper (Pitowsky 2003, 2004, 2006; Bub and Pitowsky
2010) where he puts forward his own Bayesian theory of quantum probability.
According to his analysis, the theoretical problems of QT are originated from
two ‘dogmas’: the first forbidding the use of the notion of measurement in the
fundamental axioms of the theory; the second imposing an interpretation of the
quantum state as representing a system’s objectively possessed properties and
evolution.
In this paper I illustrate and criticise Pitowsky’s analysis, taking the rejection
of the two dogmas as an accurate articulation of the fundamental stances of the
probabilistic approach to the measurement problem. In order to assess such stances,
I will test them through the use of an argument elaborated by Amit Hagar and
Meir Hemmo (2006), showing how some probabilistic interpretations of QT fail
at dictating coherent predictions in Wigner’s Friend situations.
More specifically, I evaluate three different probabilistic approaches: qBism
(more specifically, the version articulated by Christopher Fuchs, e.g. Fuchs 2010;
Fuchs et al. 2014) as a representative of the epistemic subjective interpretation of
the quantum state; Bub’s information-theoretic interpretation of QT (Bub 2016,
2018) as an example of the ontic approach to the quantum state; Pitowsky’s prob-
abilistic interpretation (Pitowsky 2003, 2004, 2006), as an epistemic but objective
interpretation. I argue that qBism succeeds in providing a formal solution to the
problem that does not lead to a self-contradictory picture, although the resulting
interpretation would bring the real subject matter of QT to clash alarmingly with
scientific practice. The other two approaches, instead, strictly fail when confronted
to scenarios where the measurement problem is relevant, showing in such a way that
they don’t provide a genuine solution to the problem.

12.2 Two Dogmas and Two Problems

The idea that deflating the quantum state’s ontological import is sufficient to explain
away the measurement problem is a leitmotif in the literature about information-
based accounts of QT, still shared by the vast majority of the advocates of the
probabilistic view:
The existence of two laws for the evolution of the state vector becomes problematical only
if it is believed that the state vector is an objective property of the system. If, however,
the state of a system is defined as a list of [experimental] propositions together with their
[probabilities of occurrence], it is not surprising that after a measurement the state must be
changed to be in accord with [any] new information. (Hartle 1968, as quoted in Fuchs 2010)
12 The Measurement Problem and Two Dogmas About Quantum Mechanics 289

Itamar Pitowsky, also together with Jeff Bub (Pitowsky 2006; Bub and Pitowsky
2010), has identified two distinct issues behind the measurement problem. The first
one, that he calls the small measurement problem, is the question “why is it hard to
observe macroscopic entanglement, and what are the conditions in which it might
be possible?” (2006, p. 28). The second issue, the big measurement problem, is the
problem of accounting for the determinateness of our experiences:
The ‘big’ measurement problem is the problem of explaining how measurements can have
definite outcomes, given the unitary dynamics of the theory: it is the problem of explaining
how individual measurement outcomes come about dynamically. (Bub and Pitowsky 2010,
p. 5)

In this paper we are exclusively concerned with the big problem; therefore,
from now on, I will refer to the latter more simply as the measurement problem.
However, I reject Bub’s and Pitowsky a priori request for a dynamical account;
therefore I characterise the measurement problem as the problem of accounting for
the determinateness of measurement results.
The second most notable contribution that Pitowsky has given to the analysis of
the measurement problem concerns the identification of two assumptions, labelled
as ‘dogmas’, in the interpretation of QT, that, according to the philosopher, are at
the origin of the issues in the interpretation of QT.
The first dogma is Bell’s assertion (defended in [(Bell 1987)]) that measurement should
never be introduced as a primitive process in a fundamental mechanical theory like classical
or quantum mechanics, but should always be open to a complete analysis, in principle, of
how the individual outcomes come about dynamically. The second dogma is the view that
the quantum state has an ontological significance analogous to the ontological significance
of the classical state as the ‘truthmaker’ for propositions about the occurrence and non-
occurrence of events, i.e., that the quantum state is a representation of physical reality.
(Pitowksy 2006, p. 5)

A couple of observations are in order here. First of all, the two claims:
“the quantum state has an ontological significance analogous to the ontological
significance of the classical state as the ‘truthmaker’ for propositions about the
occurrence and non-occurrence of events” and “the quantum state is a representation
of physical reality”, are not equivalent, although they are here put forward as such.
More specifically, one can deny that the quantum state represents ‘traditionally’
(Bub 2018) in the sense that it does not represent a system’s properties, and therefore
acting as a truthmaker for propositions about the occurrence and non-occurrence of
events, without denying that the quantum state represents physical reality. In fact,
that’s the case for Bub’s information-theoretic interpretation which, as I will argue
below, denies that the quantum state represents a system’s properties and dynamics,
and yet interprets the quantum state ontically, as representing physical reality. In
the following, therefore, I take the second dogma as the view that the quantum state
does not represent traditionally, while the issue whether the quantum state represents
physical reality or a mental state remains open.
Secondly, although probabilistic interpretations reject both dogmas, it is useful
to keep in mind that, at least at a first sight, the two are logically distinct, and it still
remains to be seen whether they need to come in package. The rejection of the first
290 L. Felline

dogma leads to a black-box interpretation, whose traditional versions put forward an

interpretation of the quantum state as representing the properties of physical systems
(e.g. Bohr’s interpretation). It might be argued that the more traditional black-box
theories have proven inconsistent, and that once one rejects the first dogma, the
explanation away of the measurement problem requires the rejection of the second.
This is, for instance, the core of Pitowsky’s and Bub’s view, according to which not
only the second dogma runs ‘very quickly’ into the measurement problem (Pitowsky
2006, p. 3), but the rejection of the latter is sufficient to solve the measurement
problem in black-box theories.
In support of this last claim, over which the critical analysis of this paper is piv-
oting, Pitowsky outlines a straightforward argument. First of all, the measurement
problem consists in the tension between the fact that measurements always yield
determinate values, and the fact that they are also performed on systems whose
states do not associate a determinate value to the measured observable. If this is so,
for the achievement of a solution it would be sufficient to reject one of the horns of
the dilemma:
The BIG problem concerns those who believe that the quantum state is a real physical
state which obeys Schrodinger’s equation in all circumstances. In this picture a physical
state in which my desk is in a superposition of being in Chicago and in Jerusalem is a real
possibility; and similarly, a superposed alive-dead cat. [ . . . ]
In our scheme quantum states are just assignments of probabilities to possible events,
that is, possible measurement outcomes. This means that the updating of the probabilities
during a measurement follows the Von Neumann-Luders projection postulate and not
Schrodinger’s dynamics. [ . . . ] So the BIG measurement problem does not arise. (Pitowsky
2006, pp. 26–27).

12.3 Incompatible Predictions in Black-Box Approaches

In this section I am going to introduce a somewhat neglected argument put forward

in Hagar (2003) and in Hagar and Hemmo (2006), which introduces a scenario
against which, according to HH, black-box interpretations of QT fail.
Let’s take an experiment with two observers, Wigner and Friend, where the latter
is inside a lab, isolated from Wigner and the rest of the environment. Before Friend
enters the lab, she and Wigner agreed upon the following protocol: in the first part
of the experiment Friend performs a z-spin measurement on a half-spin particle P in
the state:
 √ √ 
| Psi 0>P = 1 2 |+z >P + 1/ 2 |−z >P (12.1)

Therefore, the system composed of Friend and the particle (F + P) is in the state:
 √ √ 
| Psi 0>P+F = 1/ 2 | +z >P + 1/ 2 |−z >P | Psi 0>F (12.2)
12 The Measurement Problem and Two Dogmas About Quantum Mechanics 291

Since they agreed on this first part of the protocol, Wigner knows that Friend is
going to perform the measurement; however, because the lab is isolated from Wigner
and the rest of the environment, he does not know the result of her measurement.
Now, consider a second measurement, of an observable O of the system F + P,
with eigenstate
√ √
| Psi>P+F = 1/ 2 | +z >P | +z >F + 1/ 2 |−z >P |−z >F (12.3)

whose eigenvalue is, say, YES.

HH’s question is: according to a black-box approach, what predictions are
dictated by QT for the results of this measurement?
On the one hand, F + P is an isolated system; therefore, its evolution should
follow Schrodinger equation. If so, the state at the end of Friend’s measurement
should be
√ √
| Psi 1>P+F = 1/ 2 | +z >P |+z >F + 1/ 2 |−z >P |−z >F , (12.4)

which is an eigenstate of O, with eigenvalue YES. From this perspective, therefore,

QT dictates that the measurement of O will yield YES with probability 1.
On the other hand, Friend’s interaction with P is a measurement interaction,
which, according to the black-box approach, requires the application of Luders’
rule. It follows therefore that F + P’s state should collapse into either one of two
pure states

| Psi 1 > = | +z >P | +z >F (12.5)

or

| Psi 1 > = | −z >P | −z >F , (12.6)

both associating a probability ½ to the result YES in an O-measurement. According

to this line of thought, therefore, QT dictates that the probabilities for a YES result
over the measurement of O is ½.
To make things more complicated, each option adopts a specific stance about
the evolution of the quantum state in a measurement process, ruling therefore out a
black-box approach. If we take into consideration that the lab is an isolated system,
and therefore apply Schrodinger equation to the first part of the experiment, this
would correspond to a no-collapse approach. If, on the opposite, we apply Luders
rule, then we are indeed applying a collapse view of quantum measurement.
According to Hagar and Hemmo (HH), this shows that any solution to this kind
of Wigner’s friend scenario (and therefore, a fortiori, of the measurement problem)
implies an analysis of the processes behind measurement interaction. Since, they
add, every black-box interpretation of QT is equally affected by this argument,
they conclude that black-box approaches to quantum theory are unable to provide a
coherent picture of the predictions of QT.
292 L. Felline

In the rest of the paper, the application of this argument to three different
theoretical accounts will be used to explore the relationship between the two dogmas
and the measurement problem.

12.4 QBism

I have already said that HH chose qBism as a case study to formulate their argument,
while claiming that the same result applies to any black-box approach to QT. The
succeeding debate around this specific approach, however, has shed some light about
how different black-box approaches lead to different results when faced to Wigner’s
Friend scenarios. In order to formulate my criticism of qBism, in the following I
will build on such a debate. I will first isolate the exact feature of qBism that allows
it to formally escape HH’s argument, and then show how this same feature leads to
a philosophically problematic account of QT.
According to qBism, the quantum state represents the rational subjective belief
of an observer, assembled from personal experience. The first direct consequence of
this interpretation is that quantum probabilities are not a measure of anything in real-
ity, but only of the degree of belief of an observer over future experience. Secondly,
different observers might consistently attribute different quantum states to the same
system. In order to solve the dilemma in the Wigner’s Friend scenario, therefore, one
needs to take into consideration each observer’s individual perspective: given that
she has gathered new data, QT prescribes Friend to update P’s quantum state (and
F + P’s state, clearly). From her perspective, then, the prediction of the result in an
O-measurement in the second stage of the experiment will be YES with probability
1/2. On the other hand, Wigner has registered no new data, therefore QT prescribes
him to maintain the entangled state (12.4) for F + P. From his perspective, therefore,
the O-measurement will yield YES with probability 1.
The surprising consequence is that, according to qBism, different observers
can coherently hold incompatible predictions about the world. The subjectivism
embraced by qBism turns QT into “a “single-user theory”: probability assign-
ments express the beliefs of the agent who makes them, and refer to that same
agent’s expectations for her subsequent experiences. The term “single-user” does
not, however, mean that different users cannot each assign their own coherent
probabilities” (Fuchs et al. 2014, p. 2).
According to Amit Hagar (2007) this strategy can’t work: the (repeated) mea-
surement of O will yield a series of results with a certain frequency that will
(hopefully!) confirm one of the predictions and falsify the other. If this is so, at
the end of the day one of the two alternatives (collapse or no-collapse) will be right,
while the other will be wrong.
To Hagar, Timpson (2013) replies that this kind of objections are ineffective
against Qbism:
12 The Measurement Problem and Two Dogmas About Quantum Mechanics 293

[n]o objection can be successful which takes the form: ‘in such and such a situation, the
quantum Bayesian position will give rise to, or will allow as a possibility, a state assignment
which can be shown not to fit the facts’, simply because the position denies that the requisite
kind of relations between physical facts and probability assignments hold. (Timpson 2013,
p. 209)

Here is my attempt to unpack Timpson’s position: in the experiment here

illustrated, the consequence of the subjective interpretation of probabilities is that
Wigner’s beliefs about the quantum state of F + P are partially determined by a
previous subjective assignment, therefore the ‘probability 1’ of the result YES is
also merely subjective, rather than an objective certainty. Let’s say for instance
that Wigner’s experiment yields a different result from YES. According to the
ontic view of the quantum state, probabilistic predictions are determined by an
objective reality grasped by the theory. Wrong predictions, therefore, show that
there is something wrong in Wigner’s representation of reality: this is one of the
pillars of empirical sciences reasoning. However, in qBism, the relation that holds
between data and quantum state representation is not of this kind; probabilities do
not represent anything in the world, which means that Wigner’s quantum state
representation can’t be wrong about the world. As a consequence, dealing with
‘surprising’ experiences does not have to be cause of sweat for the qBist, who’s
reaction is simply to apply the same rules for updating expectations (s)he applies
every time (s)he must associate (but not discover) previously unknown chances to
an event.
However, one can still counter that, after a sufficiently large repetition of the
experiment, the falsification (on either side) of one’s predictions should lead a
rational agent to abandon the unsuccessful rules she has been following for the
formation of expectations (i.e. QT!).3 But this means, as HH argued, that QT is,
as it is, incomplete.
Another counterargument often used to block HH’s conclusion is to hinge on
the fact that there is no way for Wigner and Friend to compare their experience
and predictions: “Results objectively available for anyone’s inspection? This is the
whole issue with “Wigner’s friend” in the first place. If both agents could just “look”
at the counter simultaneously with negligible effect in principle, we would not be
having this discussion” (Fuchs and Stacey 2019, p. 9).
The argument here is that in the second measurement, P’s z-spin and Friend’s
record of it will be erased, so that any ground for Friend’s ½ predictions would
fail. However, that this is not really the core of the issue has been already argued
extensively,4 and recently explicitly illustrated by more sophisticated versions of
the Wigner’s friend thought experiment showing that the same kind of contradicting
predictions can be reproduced even in scenarios where the necessary information

3I am very grateful to Meir Hemmo and an anonymous referee for pushing me to provide a more
thorough evaluation of Timpson’s position.
4 By HH themselves first, but see also, just to cite one, (Bacciagaluppi 2013) for an insightful take

on the problematical character of the qBist solution.

294 L. Felline

is available to all the observers, while the lab is maintained in isolation (see for
instance (Frauchiger and Renner 2018) and (Baumann and Brukner 2019)).
The core of the qBist solution to this problem lies instead in “some kind of
solipsism or radical relativism, in which we care only about single individuals’
credence’s, and not about whether and how they ought to mesh” (Bacciagaluppi
2013, p. 6).
Now, it is difficult to formulate an argument with solid grounds in favour of an
interpretation of QT, if your starting point is solipsism. However, it might be argued
that the situation is not as bad as it might seem. Timpson (2013, §9.3, but see also
Healey 2017, §2), for instance, denies that qBism implies solipsism, and claims
instead that the former avoids the paradoxical or anyway pernicious consequences
of the latter as a general epistemological thesis. In other words, although, according
to qBism, QT fails to include the perspective of more than one observer, it does not
follow that according to Qbism there is only one sentient being in the world, let
alone that the rest of empirical science is affected by these limitations. This same
defence strategy is used by Timpson against the charge of instrumentalism: it is
true that, according to Qbism, QT is neither true or false of the world, and a mere
instrument for bookkeeping of probabilities. However, in the case of Qbism, this
conclusion follows, from specific considerations about the structure of QT and only
applies to said theory. Because of these considerations, qBism does not imply the
majority of the controversial features of instrumentalism as a general position about
science.
The upshot of this rebuttal to criticisms is: QT might forbid inferences about
the experiences of other observers as solipsism does, and it might be as void of
knowledge about the world, as instrumentalism dictates – but this is OK, because
qBism does not lead to the possibly paradoxical consequences of such thesis.
I am not going here to challenge Timpson’s argument. Here, I take for granted
that this defence provides qBism with a way-out from the conclusions of HH
argument. This being said, it is legitimate to scrutinize the consequences of this
unique status of QT with respect to the rest of empirical science – uniqueness that,
we just concluded, lies at the core of the qBist success.
I am going to argue that this achievement is a pyrrhic victory, as it still fails to
provide an acceptable picture of the scientific enterprise in the field of QT.
qBism succeeds in the context of Wigner’s Friend scenarios, but also in explain-
ing away other apparent oddities of the quantum world, because it characterizes QT
not as a physical theory, but as a Bayesian theory for the formation of expectations.
As such, it can coherently ignore the constraints imposed to an objective description
of the world, when they are in the way of coherence. While a theory about the world
must submit to the typical epistemic and methodological standards of empirical
sciences, Bayesian epistemology is subordinate to very different standards, for the
simple reason that different subject studies require different methodologies. The
violation of the requirement of intersubjectivity in the qBist explanation away of
non-locality (see Fuchs et al. 2014) and in the account of the Wigner’s Friend
thought experiment.
12 The Measurement Problem and Two Dogmas About Quantum Mechanics 295

In other words, Qbism plays a different game, with different rules, with respect
to physics and, more in general, with respect to natural sciences. And it escapes
strict failure when confronted against HH’s argument (as far as self-consistency is
the only requirement to avoid strict failure, but see later) only as far as QT is not
interpreted as natural science.
If this is so, a coherent and truthful endorsement of Qbism implies also the
renunciation to constraints with a great value of guidance to scientific change like
empirical adequacy, physical salience, or intersubjectivity because, in a quantum
Bayesian theory about rational belief, such criteria fail to make sense. In fact, the
renunciation to such criteria is the key for the Qbist explanation away of the oddities
of QT.
And yet, it can’t be ignored that giving up these criteria means throwing away a
huge piece of scientific practice – a hard bullet to bite for working physicists using
QT as an empirical science every day.
One might reply that this is a made-up problem: scientific practice does not, nor
should, change because of philosophical debates, so there is nothing to be worried
about. However, the unproblematic acknowledgment of such a chasm between
scientific practice and the real content of scientific theories, is an even harder bullet
to bite (if not straightforward failure) for philosophers.5
This is what one gets when submitting to a qBist view of QT. As we will see in
the next sections, those that indeed think that this is too high price to pay for internal
consistency, but still want to deny the representational role of QT, will have to do so
while maintaining in some way the empirical import of the quantum state, denied
by qBism.

5 Inhis analysis of the virtues and problems of Qbism, Chris Timpson formulates a similar
challenge to Qbism, but focusing on explanation:
It seems that we do have very many extensive and detailed explanations deriving from
quantum mechanics, yet if the quantum Bayesian view were correct, it is unclear how we
would do so. (Timpson 2013, p. 226)
Timpson’s criticism is that Qbism suffers of an explanatory deficit because it can’t explain why
QT explains; however, I think that the Qbist lack of a realist explanation for the explanatory power
of QT is hardly an unsurmountable problem.
First of all, the Qbist can reply to Timpson’s challenge that there is no reason why a theory that is
neither true nor false can’t do what false theories have done for centuries: being explanatory. If then
one insists (wrongfully, I think) that only at least approximate truth have explanatory power, then
the Qbist can simply deny a genuine explanatory power for QT: after all, the topic of explanation
in QT is a time-honoured headache in philosophy of science, and the philosophers that deny the
explanatory power of QT are not few. An objection hinging on the explanatory power of QT can
go just as far.
296 L. Felline

12.5 Bub’s Information-Theoretic Interpretation of QT

According to Jeff Bub, QT is a non-representational, probabilistic theory. Yet,

while rejecting a traditional representational interpretation of the quantum state, his
information-theoretic approach interprets the quantum state ontically, as a physical
state, a complete description of a quantum system (Bub 2016, p. 222). More
specifically, QT is a theory about information ‘in the physical sense’: a structure
of correlations between intrinsically random events, which is different from the
classical structure of correlations measured by Shannon information. Moreover,
according to Bub, information is a new kind of physical primitive, whose structure
“imposes objective pre-dynamic probabilistic constraints on correlations between
events, analogous to the way in which Minkowski space-time imposes kinematic
constraints on events” (Bub 2018, p. 5).
Bub adopts and contributes to Pitowsky’s analysis of the two dogmas (Bub and
Pitowsky 2010; Bub 2016, § 10) and in fact he also puts them at the origin of
the issues gravitating around the measurement problem. However, the application
of HH’s argument to Bub’s information-theoretic interpretation shows that the
rejection of the two dogmas is not sufficient to solve the measurement problem.
In HH’s thought experiment, in fact, QT provides contrasting instructions for the
prediction of measurement results. The qBist way out to inconsistency is relativizing
the quantum state to single observers. In Bub’s interpretation, on the other hand,
the quantum state represents an objective probabilistic structure that constrains the
system’s behaviour. This means that the probabilities codified by QT are objective,
and therefore that different observers must associate the same measurement with the
same predictions. In Bub’s account, QT is not a single-user theory: through its use,
Wigner and Friend must be able to come to agreement.
Let’s wrap up the results achieved so far. Contrarily to HH’s conclusions,
their argument does not rule out black-box interpretations of QT; however, Qbism
escapes failure by embracing an extremely subjective view of the quantum state that,
although dragging QT out of the realm of empirical sciences, allows to consistently
disattend constraints imposed to such sciences. Bub, on the other hand, puts forward
an information-theoretic account of QT which has the merit of maintaining the
empirical import not only of Hilbert space, but also of the quantum state. Regaining
the empirical import of the theory, however, means losing the possibility to exploit
the qBist solution to HH’s test, because, as any other empirical theory, it must
be empirically adequate and provide an intersubjective description of the non-
perspectival features of our experiences.
In the next section I will discuss Pitowsky’s information-theoretic interpretation,
a kind of compromise in between the extreme subjectivism of Qbism, and Bub’s
ontic approach.
It has to be said that Bub has recently addressed the issue of Wigner’s Friend
scenarios in the information-theoretic interpretation. However, since this new
contribution have several points of intersection with Pitowsky’s account, it would
be useful to discuss it also with his epistemic version of the information-theoretic
approach in mind.
12 The Measurement Problem and Two Dogmas About Quantum Mechanics 297

12.6 Pitowsky’s Information-Theoretic Interpretation

Pitowsky puts forward an epistemic interpretation of the quantum state as a state of

partial belief, in the sense of Bayesian probability. As we will see, differently from
qBism, his view preserves at least part of the empirical character of the quantum
state.
In Pitowsky’s information-theoretic approach, closed subspaces of the Hilbert
space correspond to possible states of the world, or events, while the quantum state
represents an agent’s uncertainty about it. In other words, here, as in qBism, the
quantum state is a device for bookkeeping of probabilities (Pitowsky 2006, p. 4).
Contrarily to Bub’s information-theoretic approach, the quantum state is therefore
a derived entity, and in fact a fundamental part of Pitowsky’s interpretational work
consists in the derivation of quantum probabilities from the structure of quantum
events. The derivation starts from the axioms of QT as formulated by Birkoff and
von Neumann (1936), of which the structure of quantum events is shown to be a
model. According to the resulting representation theorem, the space of events L in
quantum theory is the lattice of subspaces of a Hilbert space. The fact that, as a
consequence of Gleason’s theorem, quantum probabilities can be derivable from
this sole structure, constitutes, according to Pitowsky, ‘one of the strongest pieces
of evidence in support of the claim that the Hilbert space formalism is just a new
kind of probability theory’ (2006, p. 14).
Finally, adopting Bayesian probability as a conceptual and formal background,
means analysing probabilities through rational betting behaviour, i.e. they are
measured ‘by proposing a bet, and seeing what are the lowest odds he will accept’
(Ramsey 1926, as cited in Pitowksy 2006, p. 15).
Few clarifications are here in order.
In Pitowsky’s view, the notion of measurement appears in the axioms of QT:
QT does not make predictions about generic occurrences and correlations between
them – in this betting game, agents can only bet on measurements results, and,
between them, only those that can be considered facts, where not every measurement
result is a fact:
by “fact” I mean here, and throughout, a recorded fact, an actual outcome of a measurement.
Restricting the notion of “fact” in this way should not be understood, at this stage, as a
metaphysical thesis about reality. It is simply the concept of “fact” that is analytically related
to our notion of “event”, in the sense that only a recordable event can potentially be the
object of a gamble. (2006, p. 10)

If a measurement is performed of which no result is registered, then there is no

fact of the matter about the result of that measurement. This does not mean that
according to Pitowsky such results do not exist, but rather that they are not ‘events’
or ‘facts’ over which the theory is allowed to make predictions.
The second clarification concerns the second dogma, i.e. the reality of the
quantum state. I have said that, in Pitowsky’s view, the quantum state represents
a bookkeeping device for keeping track of probabilities, which are different from
classical probabilities. As we have seen, the risk for an epistemic views of the
298 L. Felline

quantum state run the risk of as a bookkeeping device is that it turns out being void
of physical content, or, even worse, instrumentalist; part of Pitowsky’s conceptual
work is therefore devoted to stressing how this approach avoids it.
One way to mark a distance from instrumentalism is to stress by stressing the
explanatory power of quantum Bayesianism. Under an instrumentalist view (e.g. in
the textbook approach to QT), in fact, the question ‘why do the quantum events do
not conform to classical probabilities?’ has no answer. The information-theoretic
view, on the other hand, have more tools to provide such an explanation, i.e. since
it is realist towards the structure of quantum gambles, i.e. towards the structure of
Hilbert space
the Hilbert space, or more precisely, the lattice of its closed subspaces, [is] the structure that
represents the “elements of reality” in quantum theory. (Pitowsky 2006, p. 4)
Instrumentalists often take their “raw material” to be the set of space–time events: clicks
in counters, traces in bubble chambers, dots on photographic plates, and so on. Quantum
theory imposes on this set a definite structure. Certain blips in space–time are identified as
instances of the same event. Some families of clicks in counters are assumed to have logical
relations with other families, etc. What we call reality is not just the bare set of events, it is
this set together with its structure, for all that is left without the structure is noise. [ . . . ]
It is one thing to say that the only role of quantum theory is to “predict experimental
outcome” and that different measurements are “complementary.” It is quite another thing to
provide an understanding of what it means for two experiments to be incompatible, and yet
for their possible outcomes to be related; to show how these relations imply the uncertainty
principle; and even, finally, to realize that the structure of events dictates the numerical
values of the probabilities (Gleason’s theorem). (Pitowksy 2003, p. 412)

The fact that such a structure has a physical reality allows the above-cited
explanations to be genuine physical explanations rather than collapsing to the kind
of explanations given by logic or formal epistemology. Said physical explanations
are not dynamical, but structural, in the sense of (Felline 2018a, b): the explanandum
phenomenon is explained as the instantiation of a physical structure, fundamental
in the sense that it is not inferable from the properties and behaviour of underlying
entities.
Pitowsky’s realism goes further than this. I have said before that, according to
Pitowsky’s account, only registered measurement results are facts, about which
the theory can make predictions. This claim, however, has an exception when the
quantum state also describes an objective outside reality, and its predictions can be
taken as genuine facts:
In the Bayesian approach what constitutes a possible event is dictated by Nature, and the
probability of the event represents the degree of belief we attach to its occurrence. This
distinction, however, is not sharp; what is possible is also a matter of judgment in the sense
that an event is judged impossible if it gets probability zero in all circumstances. In the
present case we deal with physical events, and what is impossible is therefore dictated by
the best available physical theory. Hence, probability considerations enter into the structure
of the set of possible events. We represent by 0 the equivalence class of all events which
our physical theory declares to be utterly impossible (never occur, and therefore always get
probability zero) and by 1 what is certain (always occur, and therefore get probability one).
(Pitowsky 2006, p. 5)
12 The Measurement Problem and Two Dogmas About Quantum Mechanics 299

So, according to this stipulations, not only registered results are facts, but also
results with probability 1. In this case, the quantum state represents something in
the world.
This traces a crucial difference between Pitowsky’s information-theoretic inter-
pretation and Qbism. In the latter, given the strict Bayesian reading of probabilities,
probability 1 only means that the person who associates this probability to an event
E very strongly beliefs the occurrence of E (the same is valid, mutatis mutandis, for
probability 0).
According to Pitowsky’s analysis of the measurement problem, adopting an
epistemic view is sufficient to explain away the measurement problem. In fact,
the claim that the quantum state is a mental state blocks, at a first analysis, the
consequences that we have seen in Bub’s ontic approach.
This, however, is not a viable solution for Pitowsky’s approach, due to the realism
he acknowledges towards the quantum state.
In order to concretely see why, take again HH’s argument and the predictions
for the measurement of O. In a partially objective reading of the state like Pitowsky
epistemic view, the fact that Wigner associates probability 1 to the result YES means
that such result is an objective fact about reality. But if this is so, then there is a fact
of the matter about what will be the result of the measurement of O and Friend’s
predictions are objectively wrong.
But this, we know, would mean adopting a non-collapse view of QT.

12.7 A Bohrian Escape

Recently (2018) Bub has formulated a new argument addressing the problem of
incompatible predictions of QT in Wigner’s Friend scenario. Bub does not directly
cite HH’s argument, and focuses instead on the more recent argument formulated in
(Frauchiger and Renner 2018). In the following, for simplicity, I will keep using the
argument as formulated by HH.
Bub stance about the notion of measurement is very much in line with Pitowsky’s
genuine facts. Accordingly, QT does not apply to, nor make predictions about,
generic events, but only about measurement results. Here, as in Pitowsky, the notion
of measurement is substantially revised:
A quantum “measurement” is a bit of a misnomer and not really the same sort of thing as a
measurement of a physical quantity of a classical system. It involves putting a microsystem,
like a photon, in a situation, say a beamsplitter or an analyzing filter, where the photon is
forced to make an intrinsically random transition recorded as one of two macroscopically
distinct alternatives in a device like a photon detector. The registration of the measurement
outcome at the Boolean macrolevel is crucial, because it is only with respect to a suitable
structure of alternative possibilities that it makes sense to talk about an event as definitely
occurring or not occurring, and this structure is a Boolean algebra. (Bub 2018, p. 6)

So, on the one hand Bub provides a physical characterization that partially
explicates the notion of measurement (it involves an intrinsically random transition
300 L. Felline

for the measurement system, whose final state must be registered macroscopically);
on the other hand, the first dogma is still violated, since there is no criterion for
whether or not a process counts as measurement (we don’t know when a photon is
forced to make an intrinsically random transition).
Bub explains the problems in Wigner’s Friend scenario as originated by the
structure of quantum events, composed by a “family of “interwined” Boolean
algebras, one for each set of commuting observables [ . . . ]. The interwinement
precludes the possibility of embedding the whole collection into one inclusive
Boolean algebra, so you can’t assign truth values consistently to the propositions
about observable values in all these Boolean algebras” (Bub 2018, p. 5).
So, the algebra of observables is non-Boolean, but each observable (each
measurement) picks up a Boolean algebra which associates different probabilities to
same observables. In our thought experiment: Friend’s and Wigner’s measurements
pick up different Boolean algebras that can’t be consistently embedded into one
single coherent framework.
According to Bub only one Boolean algebra provides the correct framework
for the description of the state. In order to justify the selection of one single
Boolean algebra, Bub starts from reminding how George Boole introduced Boolean
constraints on probability as “conditions of possible experience”. He therefore takes
a ‘Bohrian’ perspective and suggests that the choice of a unique Boolean algebra
is imposed by the necessity to “tell others what we have done and what we have
learned” (Bohr, as cited in Bub 2018, p. 9).
The correct quantum state to associate to a system, therefore, is the one
corresponding to what Bub calls the ultimate measurement (and relative ultimate
observer). It is important to notice right from the start, and I will insist on this point
later along the way, that in order for HH’s objection to be met, all observers (not
only the ultimate observer) must apply the same Boolean algebra, on pain of falling
again into a scenario with inconsistent predictions.
But who is the ultimate observer in HH’s thought experiment?
The only criterion for the identification of the ultimate measurement/observer is
that the result of the measurement must be registered macroscopically.
“In a situation, as in the [Hagar and Hemmo] argument, where there are
multiple candidate observers, there is a question as to whether [Friend is] “ultimate
observer[..],” or whether only Wigner [is]”. The difference has to do with whether
[Friend] perform measurements of the observables [P] with definite outcomes at the
Boolean macrolevel, or whether they are manipulated by Wigner [ . . . ] in unitary
transformations that entangle [Friend] with systems in their laboratories, with no
definite outcomes for the observables A and B. What actually happens to [Friend]
is different in the two situations” (Bub 2018, p. 11).
It is important to clarify, about this passage, that the exclusive character of the
disjunction ‘Friend performs measurements registered at the macrolevel or Wigner
manipulates observables’ is misleading. The two disjuncts are actually not mutually
excluding: both can be true. Let’s assume, in fact, that both Wigner and Friend write
down their results and are therefore ultimate observers. This is an unproblematic
12 The Measurement Problem and Two Dogmas About Quantum Mechanics 301

assumption, since, according to Bub’s analysis,6 both Friend and Wigner act as
ultimate observers in different moments of the experiment: in the first part, when
Friend performs her measurement, she is the ultimate observer, while Wigner
becomes the ultimate observer in the second part of the experiment.
Now, Wigner knows (as we know) that Friend will write down her measurements’
results, and so that she is the legitimate ultimate observer in the first part of the
experiment. He also knows that he has to take this assignment very seriously, given
that “What actually happens to [Friend] is different” whether or not Friend is the
ultimate observer and “the difference between the two cases [ . . . ] is an objective
fact at the macrolevel” (Bub 2018, p. 12).
This means that Wigner knows that in the first part of the experiment a legitimate
ultimate measurement is being performed in the lab, and with it the uncontrollable
disturbance that characterize, in measurement processes, the passage from non-
Booleanity to Booleanity. If we were to use Pitowsky’s Bayesian framework, we
could say that Friend’s measurement is a gamble over which Wigner can bet. He
does not clearly know what is the result of the experiment, but he knows that she
definitely gets a + or a− result, because this is what the ‘ultimate Boolean algebra’
predicts.
The point I am trying to make is that, if QT is not a ‘single user theory’ (as
it is not according to Bub), and if the selection of a Boolean algebra does not
apply uniquely to the ultimate observer (as it does not according to Bub), then
after Friend’s measurement, Both she and Wigner will have to predict that her
measurement ends up either in the state (12.5) or in the state (12.6). The result of
Friend’s experiment is an objective fact about the world, which Wigner can’t ignore.
The correct state that, according to this view, Wigner should associate to F + P
after Friend’s measurement is therefore a proper mixture of (12.5) and (12.6).
But if this is true, when Wigner makes his predictions about the measurement of
O, he should use the probabilities dictated by a proper mixture of (12.5) and (12.6),
rather than those dictated by the entangled state (12.4).
But this, again, “would seem to require a suspension of unitary evolution
in favour of an unexplained “collapse” of the quantum state” (12.3), which the
information-theoretic view clearly rejects.

6 “If there are events at the macrolevel corresponding to definite measurement outcomes for Alice
and Bob, then Alice and Bob represent “ultimate observers” and the final state of the combined
quantum coin and qubit system is |hA |0B or |tA |0B or |tA |1B , depending on the outcomes.
If Wigner and Friend subsequently measure the super-observables X, Y on the whole composite
Alice-Bob system (so they are “ultimate observers” in the subsequent scenario), the probability of
obtaining the pair of outcomes {ok, ok} is 1/4 for any of the product states |hA |0B or |tA |0B or
|tA |1B ” (Bub 2018).
302 L. Felline

12.8 Conclusions

In this paper I have tested the solutions to the big measurement problem provided
by different interpretations of QT as a theory about information, by analysing how
they behave in a thought experiment à la Wigner’s Friend.
As analytical tools for the examination of such performances, I’ve used the two
claims that, according to Itamar Pitowsky, lie at the basis of the conundrums of QT:
the claim that the concept of measurement should not appear in the fundamental
axioms of a physical theory and the claim that the quantum state represents a
physical system, its properties and its dynamical evolution.
About the first dogma, HH argue that the conclusion of their argument is that
black-box approaches are incoherent or not complete. If this is true, their argument
can be seen as a justification of the first dogma, which, as a result, is not a dogma
anymore: besides the already good reasons illustrated by John Bell, a reason to adopt
Bell’s dictum is that black-box theories can’t provide consistent predictions. In the
previous sections I have acknowledged that qBism might have a way-out of HH’s
conclusions. Should we conclude that HH’s argument fails in vindicating the first
dogma?
I think this conclusion would be wrong. Bell’s criticism of the notion of
measurement is a reflection over physical theories: measurement can’t appear in
a fundamental physical theory because it is not a physically fundamental notion. It
should be already clear that, since the conclusion of my analysis is that qBism is not
a physical theory, the first dogma does not apply to it.
When seen, as it was intended, as a criterion for fundamental physics, Bell’s
dictum seems quite reasonable. However, it is hardly surprising that the notion of
measurement, and measurement results, appear in the axioms of a theory about
beliefs and about how to update beliefs with new experience.7
The fact that Qbism is not affected by HH’s argument, therefore, does not say
much about the status of Bell’s dictum as a dogma or as a justifiable request. As far
as HH’s argument is successful against physical black-box theories, it still provides
a valid justification for the first dogma.
From the analysis put forward in this paper, therefore, Bell’s dictum seems in
great shape, and not a dogma at all.
Let’s move to the analysis of the second dogma. The main target of this paper
was the assumption, common to probabilistic and information-based approaches to
QT, that the second dogma is responsible for the measurement problem, and that its
rejection is sufficient to explain the problem away.
The failure of Bub’s and Pitowsky’s information-theoretic interpretations in the
context of HH’s thought experiment, on the one hand, and the discussed problems
in the Qbist approach, show that this assumption is too simplistic: the claim that
QT is about information does not solve the measurement problem, neither in the

7 Although (Fuchs et al. 2014, p. 2) for a reflection over the notion of measurement.
12 The Measurement Problem and Two Dogmas About Quantum Mechanics 303

physical interpretation of the notion of information (as in Bub), nor in its epistemic
interpretation (as in Pitowsky’s).

Acknowledgement The discussion of this paper at the Birmingham FraMEPhys Seminar series
and at the MCMP-Western Ontario Workshop on Computation in Scientific Theory and Practice
helped me to improve this paper. I am very thankful to the editors of this volume, and especially to
Meir Hemmo for his valuable comments, and to Veronika Baumann for the same reason.

References

Bacciagaluppi, G. (2013). A critic looks at QBism. halshs-00996289.

Baumann, V., & Brukner, Č. (2019). Wigner’s friend as a rational agent. arXiv preprint
arXiv:1901.11274.
Bell, J. S. (1987). Speakable and unspeakable in quantum mechanics. Cambridge: Cambridge
University Press.
Birkhoff, G., & von Neumann, J. (1936). The logic of quantum mechanics. Annals of Mathematics,
37, 823.
Bub, J. (2016). Bananaworld: Quantum mechanics for primates. Oxford: Oxford University Press.
Bub, J. (2018). In defense of a “single-world” interpretation of quantum mechanics. Studies in
History and Philosophy of Science Part B: Studies in History and Philosophy of Modern
Physics. arxiv.org/abs/1804.03267v1.
Bub, J., & Pitowsky, I. (2010). Two dogmas about quantum mechanics. In Many worlds (pp. 433–
459). Oxford: Oxford University Press.
Felline, L. (2018a). Quantum theory is not only about information. Studies in History and
Philosophy of Science Part B: Studies in History and Philosophy of Modern Physics.
arXiv:1806.05323.
Felline, L. (2018b). Mechanisms meet structural explanation. Synthese, 195(1), 99–114.
Frauchiger, D., & Renner, R. (2018). Single-world interpretations of quantum mechanics cannot
be self-consistent. arXiv eprint quant-ph/1604.0742.
Fuchs, C. A. (2010). QBism, the perimeter of quantum Bayesianism. arXiv preprint
arXiv:1003.5209.
Fuchs, C. A., & Stacey, B. C. (2019, January). QBism: Quantum theory as a hero’s handbook.
Proceedings of the International School of Physics “Enrico Fermi”, 197, 133–202. arXiv
preprint arXiv:1612.07308.
Fuchs, C. A., Mermin, N. D., & Schack, R. (2014). An introduction to QBism with an application
to the locality of quantum mechanics. American Journal of Physics, 82(8), 749–754.
Hagar, A. (2003). A philosopher looks at quantum information theory. Philosophy of Science,
70(4), 752–775.
Hagar, A. (2007). Experimental metaphysics2: The double standard in the quantum-information
approach to the foundations of quantum theory. Studies in History and Philosophy of Science
Part B: Studies in History and Philosophy of Modern Physics, 38(4), 906–919.
Hagar, A., & Hemmo, M. (2006). Explaining the unobserved—Why quantum mechanics ain’t only
about information. Foundations of Physics, 36(9), 1295–1324.
Healey, R. (2017). Quantum-Bayesian and pragmatist views of quantum theory. In E. N.
Zalta (Ed.), The Stanford encyclopedia of philosophy (Spring 2017 Edition). https://
plato.stanford.edu/archives/spr2017/entries/quantum-bayesian/
Maudlin, T. (1997). Quantum non-locality and relativity: Metaphysical intimations of modern
physics. Malden: Wiley-Blackwell.
304 L. Felline

Myrvold, W. (2018). Philosophical issues in quantum theory. In E. N. Zalta (Ed.), The Stanford
encyclopedia of philosophy (Fall 2018 edition). URL: https://plato.stanford.edu/archives/
fall2018/entries/qt-issues/
Pitowsky, I. (2003). Betting on the outcomes of measurements: A Bayesian theory of quantum
probability. Studies in History and Philosophy of Science Part B: Studies in History and
Philosophy of Modern Physics, 34(3), 395–414.
Pitowsky, I. (2004). Macroscopic objects in quantum mechanics: A combinatorial approach.
Physical Review A, 70(2), 022103.
Pitowsky, I. (2006). Quantum mechanics as a theory of probability. In Physical theory and its
interpretation (pp. 213–240). Dordrecht: Springer. arXiv:quant-ph/0510095v1.
Ramsey, F. P. (1926). Truth and probability. Reprinted In D. H. Mellor (Ed.) F. P. Ramsey:
Philosophical Papers. Cambridge: Cambridge University Press (1990).
Timpson, C. G. (2013). Quantum information theory and the foundations of quantum mechanics.
Oxford: Oxford University Press.
Wallace, D. (2016). What is orthodox quantum mechanics? arXiv preprint arXiv:1604.05973.
Chapter 13
There Is More Than One Way to Skin
a Cat: Quantum Information Principles
in a Finite World

Amit Hagar

Abstract An analysis of two routes through which one may disentangle a quantum
system from a measuring apparatus reveals how the no–cloning theorem can follow
from an assumption on an infrared and ultraviolet cutoffs of energy in physical
interactions.

Keywords Finite nature hypothesis · Quantum information · No cloning · No

signaling · Quantum Zeno effect · Quantum adiabatic theorem

13.1 In Memory

Itamar Pitowsky was my teacher and my MA thesis advisor. He was the reason
I became an academic. His classes were a sheer joy, and his deep voice and
enthusiasm gave warmth to us all in those dreary Jerusalem winters. Itamar’s genius
became evident a decade earlier when he published his Ph.D. thesis in the Physical
Review (Pitowsky 1983) and later in a Springer book (Pitowsky 1989) that has
gone missing from so many university libraries because everybody steals it. In it
he showed that one could still maintain a local realistic interpretation of quantum
mechanical correlations if one would allow the existence of non-measurable sets.
In a sense it was the beginning of his view of quantum mechanics as a theory of
(non–Boolean) probability, a view which he has been advocating forcefully later in

I am thankful to G. Ortiz for discussion, to the audience in the IU Logic Colloquium and the UBC
Physics Colloquium for their comments and questions on earlier versions of this paper, and to
Itamar, the oracle from Givat Ram, whose advice has guided me since I met him for the first time
more than 25 years ago.

A. Hagar ()
HPSC Department, Indiana University, Bloomington, IN, USA
e-mail: hagara@indiana.edu

© Springer Nature Switzerland AG 2020 305

M. Hemmo, O. Shenker (eds.), Quantum, Probability, Logic,
Jerusalem Studies in Philosophy and History of Science,
https://doi.org/10.1007/978-3-030-34316-3_13
306 A. Hagar

his career before his untimely death. I dedicate this paper to Itamar and his unique
empiricist view. The calm sunshine of his mind is among the things I will never
forget.

13.2 Introduction

In the late 1920s von Neumann axiomatized non relativistic quantum mechan-
ics (QM henceforth) with the Hilbert space, and ever since then generations
of physicists have been accustomed to view this space as indispensable to the
description of quantum phenomena. The Hilbert space with its inner product and
the accompanying non commutative algebra of its orthomodular lattice of closed
subspaces have also given rise to a manifold of “interpretations” of non relativistic
QM, and to a subsequent plethora of puzzles and metaphysical speculations about
the inhabitant of this space, the “state vector” (Nye and Albert 2013), its “branching”
(Saunders et al. 2010), and its “collapse” (Ghirardi et al. 1986).
Such puzzles are not surprising. Faithful as they are to the Galilean imperative
(“The book [of Nature] is written in mathematical language” Galilei 1623/1960),
theoretical physicists often incline to interpret mathematical constructs that they
use in the description of physical reality as designating pieces of that reality.
In classical mechanics such a correspondence is relatively straightforward: the
dynamical variables of the theory can be assigned values, and so can the attributes of
the physical objects they represent. In QM, however, things are more complicated.
To begin with, the mathematical structure can support two types for dynamical
variables, the operators and the vectors they operate on (the historical reason for this
duality can be traced back to the birth of quantum theory; see, e.g., Beller (1999,
pp. 67–78)). Next, this mathematical structure doesn’t yield itself to correspondence
with physical reality as easily as in classical mechanics. One candidate for the
dynamical variables, the operator, can be assigned (eigen)values only in specific
circumstances (namely, when it operates on eigenvectors), and, furthermore, not
all operators can be assigned (eigen)values simultaneously (Kochen and Specker
1967); the other candidate, the state vector, can be assigned values (amplitudes),
but instead of actual attributes of a physical object, these (or rather their mod
square) represent the probabilities of the object’s having these attributes as an object
represented by (or spanned by the eigenvectors of) the respective operator.
Those who insist on interpreting the state vector of a single quantum system
as designating a real physical object can do so by (i) altering QM, namely by
introducing a physical collapse of the state vector as a genuine physical process
(Ghirardi et al. 1986), or by (ii) replacing QM with a deterministic and non–local
alternative (Goldstein et al. 1992), or by (iii) keeping QM intact but committing to
a metaphysical picture of multiple worlds (Saunders et al. 2010). Since neither of
these routes is terribly attractive to physicists, a possible alternative is to regard the
state vector as a mathematical object which does not represent pieces of reality, but
rather our knowledge of that reality (Unruh 1994), knowledge which is inherently
13 There Is More Than One Way to Skin a Cat: Quantum Information. . . 307

incomplete (Caves et al. 2002). On this view, QM is regarded a theory of probability

(Bub and Pitowsky 2010), with a non–Boolean structure that allows the physicist to
make bets on the results of experiments.
The idea of QM as a theory of probability (and of the Hilbert space as a probabil-
ity space) allows us to embark on all sorts of projects that, for example, aim to derive
the structure of the Hilbert space from basic principles of (quantum) information
(Clifton et al. 2003), or reconstruct QM as a theory of information (Fuchs 2003).
In these projects interference and entanglement become manifestations of the non–
Boolean structure of quantum probabilities, a structure that allows us to perform
dense coding, teleportation, and cryptography and prevents us from signaling or
cloning. One common feature of these projects is that they all take measurement
results as primitives, and refrain from dynamically analyzing further how these
results come about.
This strategy has gained a lot of support in the quantum foundations community
(to the extent that quantum foundations itself has become by and large a branch of
quantum information science). But no matter how attractive you find it, portraying
QM as a theory of statistical inference still leaves us with an unresolved tension:
given that on this view the probabilities that are employed in theory refer to our
knowledge, then one cannot argue anymore that QM is a theory about the world;
it is, rather, a theory about our knowledge. Contrast that with the advertised role of
physics to describe the natural world (and not our state of mind while observing it).
Elsewhere I have argued for a way to ease this tension by interpreting “lack of
knowledge” objectively as a natural limit on spatial resolution that can be traced
to the finite nature of space (Hagar and Sergioli 2014; Hagar 2014, 2017). Here I
would like to show how other features of the information–theoretic approach to QM
can arise within such a finitist picture.
I will focus on the famous no–cloning theorem that says that an unknown
quantum state cannot be cloned, and show how this information–theoretic principle
can emerge from an examination of the limitations that exist on disentangling of
a quantum system from its measurement apparatus in two generic scenarios, the
quantum adiabatic theorem and the quantum Zeno effect.
After briefly introducing the no–cloning theorem and the finite nature hypothesis,
in Sects. 13.3.1 and 13.3.2 I will show that in both scenarios there is no way to keep
a quantum state undisturbed during a measurement process unless prior knowledge
thereof exists. What prevents us from doing so in both cases, I shall demonstrate, is
a finitist assumption on interaction energy. In this sense, I shall argue in Sect. 13.4,
the no–cloning theorem can be seen as a result of the finite nature hypothesis.

13.3 Knowledge, or Lack Thereof

The famous eigenvector/eigenvalue rule in QM states that a quantum system

possesses a definite value for a specific property when it is an eigenstate of the
operator that represents that property. Now we certainly know how to prepare a
308 A. Hagar

system so that it will end up in a specific state, i.e., we measure the system in a
certain orthonormal basis of an operator and “collapse” it to one of the eigenvectors
of the operator. Note that the quantum state we prepared becomes “known” to us
with certainty only after the measurement took place, and once it is, we can repeat
the measurement many times without changing it.
That we need to accept this scenario and not analyze the measurement process
any further is the essence of the information–theoretic view that regards measure-
ment results as primitive and “collapse” as a subjective knowledge update and not as
a physical process. But what if we asked the reverse question? how does one prepare
an operator so that an unknown quantum state becomes its eigenstate?
The obstacle for doing so comes in the form of the no–cloning theorem, that
states that a cloning device of the sort |ψ ⊗ |s, where |ψ is an unknown quantum
state and |s is the pure target state to which |ψ should be cloned according to the
unitary evolution

|ψ ⊗ |s →U U (|ψ ⊗ |s) = |ψ ⊗ |ψ (13.1)

can only work if |s = |ψ (in which case |ψ is known in advance) or if |ψ and |s
are orthogonal. In other words, a general unknown quantum state cannot be cloned
with unitary evolution (Dieks 1982; Wootters and Zurek 1982).
From an information–theoretic perspective, the theorem encapsulates the differ-
ence between quantum and classical information: in the classical case states such
as |0 and |1 are always orthogonal but QM allows for additional non orthogonal
states such as √1 (|0 + |1). From an operational perspective, however, the no–
2
cloning theorem is just another reminder that in QM the measurement process of
an unknown state disturbs the state (Park 1970).1 It is this feature that, or so I
shall argue, the finite nature hypothesis can explain in the two generic measurement
scenarios described below.
The finite nature hypothesis imposes an upper bound on energy in any physical
interaction. It translates into physics a position in the philosophy of mathematics
known as finitism, according to which the physical world is seen as “a large but
finite system; finite in the amount of information in any volume of spacetime, and
finite in the total volume of spacetime” (Fredkin 1990). On this view, space and
time are only finitely extended and divisible, all physical quantities are discrete,
and all physical processes are only finitely complex. What follows is an attempt to
demonstrate how the operational principle behind the no–cloning theorem results
from such a bound on energy in the interaction between the quantum system and its
measuring device.

1 If
we have many identical copies of a qubit then it is possible to measure the mean value of non–
commuting observables to completely determine the density matrix of the qubit Inherent in the
conclusion that nonorthogonal states cannot be distinguished without disturbing them then is the
implicit provision that it is not possible to make a perfect copy of a qubit.
13 There Is More Than One Way to Skin a Cat: Quantum Information. . . 309

13.3.1 How Slow Is Slow Enough

One way a construction of a measurement interaction that keeps the quantum state
intact could take place is via the quantum adiabatic theorem. According to this
theorem (Messiah 1961, pp. 739–746), an eigenstate (and in particular a ground
state) of an initial time–dependent Hamiltonian of a quantum system can remain
unperturbed in the process of deforming that Hamiltonian, if several conditions are
met: (i) the initial and final Hamiltonians, as well as the ones the system is deformed
through, are all non–degenerate, and their eigenvalues are piecewise differentiable
in time (ii) there is no level–crossing (i.e., there always exists an energy gap between
the ground state and the next excited state), and (iii) the deformation is adiabatic,
i.e., infinitely slow. To take an everyday analogy, if your sound sleeper baby is sound
asleep in her cradle at the living room, then moving the cradle as gently as possible
from that room to the bedroom will not wake her up.
More precisely, consider a quantum system described in a Hilbert space H by a
smoothly time–dependent Hamiltonian, that varies between HI and HF , for t ranging
over [t0 , t1 ], with a total evolution time given by T = t1 − t0 , and with the following
convolution:
 
t t
H = H (t) = 1 − HI + HF . (13.2)
T T

When the above conditions (i–iii) of the adiabatic theorem are satisfied, then if the
system is initially (in t0 ) in an eigenstate of the initial Hamiltonian HI , it will evolve
at t1 (up to a phase) to an eigenstate of HF . In the special case where the eigenstate is
the ground state, the adiabatic theorem ensures that in the limit T → ∞ the system
will remain in the ground state throughout its time–evolution.
Now, although in practice T is always finite, the more it satisfies the minimum
energy gap condition (condition (ii) above), the smaller will be the probability that
the system will deviate from the ground state. This means that the gap condition
controls the evolution time of the process, in the exact following way:

gmin = min (E1 (t) − E0 (t)). (13.3)

0≤t≤T

T ! 1/gmin
2
. (13.4)

where E0 and E1 are the ground state and the first excited state, respectively.
Prima facia the adiabatic theorem seems highly suitable for a measurement
interaction that keeps the unknown quantum state intact, as it appears to allow one to
remain agnostic about both the final state vector at t1 and the final Hamiltonian HF ,
while ensuring that the former is an eigenstate (and, in particular, a ground state, in
the case one starts at t0 from a state one knows to be the ground state of HI ) of the
latter.
310 A. Hagar

One may already point out that the adiabatic theorem is correct only at the limit
T → ∞. In actual physical scenarios, T is always finite, and so there is always a
finite probability that the system will not remain in the ground state after all. In those
cases, repeated measurements of the same state would send it to a state orthogonal
thereto, and so one would not be able to fully characterize it.
However, here I shall argue that the issue isn’t just with the impracticality of
the above scenario, and the problem is actually much more serious: for even for
a finite T , one can disentangle the system from the apparatus and keep the state
undisturbed during the measurement interaction if and only if one knows in advance
the Hamiltonian, which means that one also knows in advance its ground state.
As we shall see, that T is always finite and that prior knowledge of the
Hamiltonian is required are two facets of the same fact, namely, that the question
whether a state remains undisturbed during a process of the kind described above
is an undecidable question. One could argue that there “exists” such a situation in
which the state is undisturbed by the process, and that one can make the probability
of failure of the adiabatic process as small as one wants by simply enlarging the
finite evolution time T . The problem, however, is that even in the realistic case of
a finite T , without prior knowledge of the Hamiltonian, the question what T is (i.e.,
when to halt the adiabatic process so that the state one ends up with is undisturbed),
is undecidable; it is simply impossible to compute T in advance, unless one has
prior knowledge of the final Hamiltonian (and thus of the ground state that can be
calculated therefrom).
To see why, let us focus on the gap condition (Eq. 13.4). This condition is in the
heart of the adiabatic theorem, as it ensures its validity, and, more important, as it
governs the evolution time of the entire process. This spectral gap g = E1 − E0
may be calculated for HI at t0 , since we ourselves construct this Hamiltonian and
prepare the system to be in its ground state. But how does this gap behave along
the convolution? Note that the question here is not whether the gap exists along
the convolution. Saying that T is finite means that at some point a level–crossing
would occur with a finite probability, and so the process would not keep the state
undisturbed forever. But let’s concede that a gap (hence an undisturbed state) does
“exist”. Conceding this, the question we ask now is how small is this gap in all
instances other than t0 . For in the absence of a detailed spectral analysis, in general
nobody knows what g is, how it behaves, or how to compute it!
This question of the gap behavior is of crucial importance. Recall that the gap
controls the velocity of the process with which one would like to keep the state
vector undisturbed throughout the measurement process. To achieve this, one needs
to deform the initial known Hamiltonian HI slow enough, until one arrives at the
final unknown Hamiltonian HF . But how slow is slow enough? How slow should
one go, and, more important, when should one stop the process if one doesn’t
know in advance either (a) the spectral gap or (b) the final Hamiltonian HF ? The
problem, therefore, is not that the evolution time T is finite, but that since g or HF are
unknown, T (and hence the “speed” T −1 ), are finite alright, but also unboundedly
large (slow).
13 There Is More Than One Way to Skin a Cat: Quantum Information. . . 311

Can we estimate T in advance? For each given “run” of the adiabatic process that
ends with HF we have to come up with a process whose rate of change is ∼ Tf−1 .
How do we know that we are implementing the correct rate of change while H (t)
is evolving? Apparently, by being able to measure differences of order Tf−1 , that is,
having the sensitive “speedometer” at our disposal. But when the going gets tough,
we approach very slow speeds of the order of ∼ Tf−1 , which begs the question,
since we can then compute Tf using our “speedometer”, and hence compute HF
and determine its ground state without even running the adiabatic process.
Now if we don’t have a “speedometer”, then even if we decided to increase the
running–time from the order of Tf to, say, the order of Tf + 7, we will have no clue
that the machine is indeed doing (Tf + 7)−1 and not Tf−1 , and that the state remains
undisturbed. And so in the absence of knowledge in advance of g or HF , we will
never know how slow should we evolve our physical system. But then we will also
fail to fulfill the adiabatic condition which ensures us that in any given moment the
system remains in its ground state.

13.3.2 How Fast Is Fast Enough

Another route to realizing a scenario in which a measured state remains undisturbed

is the well–known quantum Zeno effect (Degasperis et al. 1974; Peres 1980), also
known as “the watched pot” effect. Here the idea is that the survival probability of
an evolving quantum state is predicted to be altered by a process of repeated obser-
vations with a macroscopic apparatus. In the ideal case of continuous measurement
process, the state remains “frozen” and doesn’t change in time. In the more realistic
case of a finite measurement process, the probability that the state changes vanishes
as the frequency of the repeated measurements increases. In effect we have here a
mirror image of the quantum adiabatic theorem: there, one could keep the (ground)
state intact by interacting with it infinitely slow; here one can keep the (excited)
state intact by interacting with it infinitely fast.
More precisely, consider a quantum system composed of an unknown state |ψ
which interacts with a measuring apparatus, and evolves according to the total time–
independent Hamiltonian H (units are chosen so that h̄ = 1). The probability of the
“survival” of the initial state is

|(ψ|e−iH t |ψ)|2 ≈ 1 − (H )2 t 2 − . . . , (13.5)

where H is the standard deviation defined as

(H )2 = H ψ|H ψ − (ψ|H ψ)2 . (13.6)

It follows that if the projection operator on the initial state ψ is measured after a
short time t, the probability of finding the state unchanged is
312 A. Hagar

p = 1 − (H )2 t 2 , (13.7)

but if the measurement is performed n times, at intervals t/n, there is a probability

p = [1 − (H )2 (t/n)2 ]n > 1 − (H )2 t 2 , (13.8)

that in all measurements one will find the state unchanged. In fact, for n → ∞,
the left hand side in (13.8) tends to 1, and so in the limit when the frequency of
measurements tends to infinity, the state will not change at all, and will remain
“frozen”.
From a mathematical perspective, the situation can be explained as a case in
which, given a certain orthonormal basis for H = H0 + V , all the diagonal terms
reside in H0 , and V contains only off–diagonal elements. The eigenspace of H0
is thus a “decoherence free subspace” (Lidar and Whaley 2003), a sector of the
system’s Hilbert space where the system is decoupled from the measuring apparatus
(or more generally, from the environment) and thus its evolution is completely
unitary. Physically this means that the additional term V in the Hamiltonian H
rapidly alters the phases of the eigenvalues, and the decay products (responsible
for the decaying of the state) acquire energies different from those of H , as to make
the decay process more difficult, or, in the limit, impossible. Now if one could locate
one’s state vector in the eigenspace of H0 , then consecutive measurements would
not change the state vector, and one could measure it without disturbing it, and
violate the no cloning theorem.
The quantum Zeno effect was originally presented in terms of a paradox (Misra
and Sudarshan 1977), the idea being that without any restrictions on the frequency
of measurements (time being what it is, namely, a continuous parameter), QM fails
to supply us predictions for the decay of a state which is monitored by a continuous
measurement process, and so appears to be incomplete.
But as in the earlier case of the adiabatic theorem, the state can be guaranteed
to undisturbed in the context of the quantum Zeno effect only at the limit n → ∞
(i.e., T → 0). In actual physical scenarios, T = 0 hence the frequency of the
measurement is always finite, and so there is always a possibility that the state
will be disturbed after all. As in the case of the adiabatic theorem, one could just
replace actual infinity with potential unboundedness, and argue that a decoherence
free subspace simply “exists”, and, moreover, that one can make the probability of
decay as small as one wants.
And yet, such a decoherence free subspace – a subspace in which one can
measure a quantum state without disturbing it – can be constructed with certainty
only if one knows in advance the state one wants to protect. Indeed, further analysis
has revealed (Ghirardi et al. 1979) that a correct interpretation of the time–energy
uncertainty relation (i) relates the measurement frequency to the uncertainty in
energy of the measuring apparatus:

T E ≥ 1 (13.9)
13 There Is More Than One Way to Skin a Cat: Quantum Information. . . 313

where T is the time duration of the measurement and E is the energy uncertainty,
and (ii) controls, via this uncertainty, the rate of decay of the state:
π
|ψ 0 |ψ t |2 ≥ cos2 (Et), for 0 ≤ t ≤ . (13.10)
2E
Before we go on to analyze these results, two remarks are in place. First,
inequality (13.10) was first derived by Mandelstam and Tamm in 1945 for E =
H , i.e., for the case where the uncertainty (spread) in energy is given in terms
of the standard deviation, under the assumption that the latter is finite. Recently it
was shown that since H can diverge, a more suitable measure of uncertainty is
required (Uffink 1993). To our purpose here this caveat is immaterial; even after
a correct measure of uncertainty is chosen, the probability of decay of the state
still depends on this measure of uncertainty. Second, while the duration of the
measurement is a parameter and not a quantum operator, the application of the time–
energy uncertainty relation is legitimate here because we are dealing with what may
be called a static experiment (Aharonov and Bohm 1961): in this case T is not a
preassigned or externally imposed parameter of the apparatus, but is on the contrary
determined dynamically through the workings of the Schrödinger equation which
describes the evolution of the wave packet through the measuring system. The wave
packet may perhaps be said to interact with the apparatus during the time interval
T and in this sense the time–energy uncertainty relation is valid, and constrains the
precision of the energy of the interaction. Note that the time–energy uncertainty
relation all by itself doesn’t constrain T to be finite. It is the additional assumption
that in realistic interaction scenarios energy is finite and bounded from above that
does the trick. See, e.g., Hilgevoord (1998, p. 399).
With this preliminaries we can now state our negative result, which is similar to
the one we have established in the case of the adiabatic theorem. For even in the
case where the frequency T −1 is finite but unbounded (hence T is not exactly 0
but arbitrarily close to 0), one still needs to know the state in advance in order to
know how fast one needs to repeat the measurements that project the state onto its
protected subspace and thereby disentangle it from the measuring apparatus.
The reason is that also here, for each given “Zeno effect” process k, we have to
come up with a repeated measurement process whose frequency is ∼ Tk−1 . But how
do we know that we are implementing the correct frequency? Apparently, by being
able to measure differences of order Tk−1 , that is, having the sensitive “speedometer”
at our disposal. But when the going gets tough we approach very fast speeds of the
order of ∼ Tk−1 , which begs the question, since we can then compute Tk using our
“speedometer”.
Now if we don’t have a “speedometer”, then even if we decided to increase the
frequency of the repeated measurements from the order of Tk−1 to, say, the order of
(Tk − 7)−1 , we will have no clue that the apparatus is indeed working at (Tk − 7)−1
and not Tk−1 . And so also here, in the absence of knowledge in advance of the
energy spread E or the desired state (from which, along with the initial state, the
energy spread can be calculated), we will never know how fast should we measure
314 A. Hagar

our physical system so that this state is protected. But then we will also fail to fulfill
the conditions for the quantum Zeno effect which ensure us that the system remains
undisturbed in its initial state.

13.4 Moral

In both scenarios presented above, the adiabatic evolution or the Zeno effect, the
issue at stake is not the practical inability to realize actual infinity; in both cases,
the upper and lower bounds on the duration of the measurement, while a practical
necessity, signify a much deeper restriction.
What our analysis in section (3) has shown is that for the state to remain
undisturbed, the decision when to halt the adiabatic process, or how fast to measure
the state in the Zeno effect scenario requires a–priori knowledge of that state. But
this means that the imposition of some upper or lower bound on measurement
duration, ultraviolet or infrared cutoffs which signify the inability to resolve energy
differences unboundedly small, or to apply unbounded energy with increasingly
frequent measurements, respectively, is tantamount to knowledge in advance of
the state to be measured. The actual duration or frequency in the two processes,
respectively, yield the precision with which one can describe the desired known
state.
In other words, if one had such unbounded resolution power or unbounded
energy, one would be able to measure an unknown quantum state without disturbing
it and discern between two non–orthogonal quantum states. Conversely, at least in
the two cases discussed here, if one would like to measure an unknown quantum
state without disturbing it, one must avail oneself to such unbounded resolution
power or unbounded energy (this, of course, need not hold in general, as we
only investigated two possible measurement scenarios, and there may be other
scenarios in some domains which may allow one to circumvent the no–cloning
theorem without availing oneself to unbounded energy. The sheer existence of those
scenarios would entail the inapplicability of QM in those domains).2 The upshot, in
short, is that if the finite nature hypothesis holds, so does the no–cloning theorem.

2 Note that even if one had bounded but large resolution one could not measure an unknown
quantum state without disturbing it. Saying that such disturbance could be “very small” misses
the point, because one would still need to know in advance what the result is supposed to be in
order to say how small is “very small” disturbance. Regardless, even a small disturbance will be
enough to validate no cloning. Indeed, Atia and Aharonov (2017) have recently shown – and I
thank a referee for this reference – that only when a Hamiltonian or its eigenvalues are known,
then there is a tradeoff between accuracy and efficiency of the kind envisioned in the “small
disturbance” remark. Their result which connects quantum information to fundamental questions
about precision measurements emphasizes a prior point made by Aharonov et al. (2002), and which
is repeated here: once a-priori knowledge of the Hamiltonian or the state is allowed, the game has
changed and one can bypass certain structural restrictions imposed by quantum mechanics. The
whole point of this chapter is that these restrictions can arise from fundamental limitations on
spatial resolution.
13 There Is More Than One Way to Skin a Cat: Quantum Information. . . 315

Can we give other quantum information–theoretic principles the same treatment?

I believe we can, at least to the no–signaling principle which can be violated if
one allows infinite precision in spatial resolution (a-la Bohm, see Hemmo and
Shenker 2013). As for the no bit commitment principle, the answer is more involved.
This third principle is different than the other two in that it actually constrains
the quantum dynamics rather than the kinematics to certify that macroscopic
superpositions are stable over time. Since we haven’t yet observed such states, I
see no reason to embark on a project to derive it from a more fundamental physical
structure.
And the moral? Beginning with his Ph.D. thesis, Itamar has always traced
the difference between the quantum and the classical back to the nature of the
probability space that underlies the two descriptions. Here and elsewhere I have
suggested that there may be a deeper reason for the non–Boolean nature of
this probability space that underlies quantum non–commutativity, measurement
disturbance, and, as argued here, no–cloning, namely, lower and upper bounds on
energy in any physical interaction. Such a view also entails that the standard Hilbert
space formalism is nothing but an ideal mathematical structure that describes an
underlying physical reality, and that there is no point in attaching to this structure
any physical significance.
We have arrived, for completely different reasons, to a similar conclusion as
the proponents of the information–theoretic view of QM, namely, that the variety
of discussions on “the reality of the wave function” or its “collapse”, on “many
worlds” or “branching”, discussions which are all predicated on the applicability
of the Hilbert space formalism to subatomic phenomena, and which have saturated
the philosophy of QM since its inception, can be regarded as merely academic.
Our moral, however, is that one should take a step further, and inquire about the
underlying physics (and not only the information–theoretic principles) behind this
formalism; physics which I have argued, is ultimately finite and discrete.
I would like to believe that such a moral would have resonated well with Itamar’s
empiricist mind.

References

Aharonov, Y., & Bohm, D. (1961). Time in the quantum theory and the uncertainty relation for
time and energy. Physical Review, 122, 1649–1658.
Aharonov, Y., Massar, S., & Popescu, S. (2002). Measuring energy, estimating hamiltonians, and
the time-energy uncertainty relation. Physical Review A, 66, 052107.
Atia, Y., & Aharonov, D. (2017). Fast-forwarding of hamiltonians and exponentially precise
measurements. Nature Communications, 8(1), 1572.
Beller, M. (1999). Quantum dialogues. Chicago: University of Chicago Press.
Bub, J., & Pitowsky, I. (2010). Two dogmas about quantum mechanics. In S. Saunders et al. (Eds.),
Many worlds? Oxford: Oxford University Press.
Caves, C., Fuchs, C., & Schack, R. (2002). Quantum probabilities as bayesian probabilities.
Physical Review A, 65, 022305.
316 A. Hagar

Clifton, R., Bub, J., & Halverson, H. (2003). Characterizing quantum theory in terms of
information-theoretic constraints. Foundations of Physics, 33, 1561.
Degasperis, A., Fonda, L., & Ghirardi, G. C. (1974). Does the lifetime of an unstable system
depend on the measuring apparatus? Nuovo Cimento, 21A, 471–484.
Dieks, D. (1982). Communication by EPR devices. Physics Letters A, 92, 271–272.
Fredkin, E. (1990). Digital mechanics: An informational process based on reversible universal
cellular automata. Physica D, 45, 254–270.
Fuchs, C. (2003). Quantum mechanics as quantum information, mostly. Journal of Modern Optics,
50(6–7), 987–1023.
Galilei, G. (1623/1960). The assayer (Il Saggiatore). In The controversy on the comets of 1618.
Philadelphia: University of Pennsylvania Press. English trans. S. Drake & C. D. O’Malley.
Ghirardi, G., Omero, C., Weber, T., & Rimini, A. (1979). Small-time behavior of quantum
nondecay probability and zeno’s paradox in quantum mechanics. Nuovo Cimento, 52A, 421–
442.
Ghirardi, G., Rimini, A., & Weber, T. (1986). Unified dynamics for microscopic and macroscopic
systems. Physical Review D, 34(2), 470–491.
Goldstein, S., Dürr, D., & Zanghi, N. (1992). Quantum equilibrium and the origin of absolute
uncertainty. The Journal of Statistical Physics, 67(5/6). 843–907.
Hagar, A. (2014). Discrete or continuous? Cambridge: Cambridge University Press.
Hagar, A. (2017). On the tension between ontology and epistemology in quantum probabilities. In
O. Lombardi et al. (Eds.), What is quantum information? Cambridge: Cambridge University
Press.
Hagar, A., & Sergioli, G. (2014). Counting steps: a new approach to objective probability in
physics. Epistemologia, 37(2), 262–275.
Hemmo, M., & Shenker, O. (2013). Probability zero in Bohm’s theory. Philosophy of Science,
80(5), 1148–1158.
Hilgevoord, J. (1998). The uncertainty principle for energy and time. II. American Journal of
Physics, 66(5), 396–402.
Kochen, S., & Specker, E. (1967). The problem of hidden variables in quantum mechanics. Journal
of Mathematics and Mechanics, 17, 59–87.
Lidar, D., & Whaley, B. (2003). Decoherence–free subspaces and subsystems. In F. Benatti &
R. Floreanini (Eds.), Irreversible quantum dynamics (Lecture notes in physics, Vol. 622, pp. 83–
120). Berlin: Springer.
Messiah, A. (1961). Quantum mechanics (Vol. II). New York: Interscience Publishers.
Misra, B., & Sudarshan, E. C. G. (1977). The zeno paradox in quantum theory. Journal of
Mathematical Physics, 18(4), 756–763.
Nye, A., & Albert, D., eds. (2013). The wave function. Oxford: Oxford University Press.
J. Park (1970). The concept of transition in quantum mechanics. Foundations of Physics, 1(1),
23–33.
Peres, A. (1980). Zeno paradox in quantum theory. American Journal of Physics, 48(11), 931–932.
Pitowsky, I. (1983). Deterministic model of spin and statistics. Physical Review D, 27, 2316.
Pitowsky, I. (1989). Quantum probability, quantum logic (Lectures note in physics, Vol. 321).
Springer Berlin Heidelberg.
Saunders, S., Barrett, J., Kent, A., & Wallace, D. (Eds.). (2010). Many worlds? Oxford: Oxford
University Press.
Uffink, J. (1993). The rate of evolution of a quantum state. American Journal of Physics, 61(10),
935–936.
Unruh, W. G. (1994). Reality and measurement of the wave function. Physical Review A, 50, 882–
887.
Wootters, W. K., & Zurek, W. (1982). A single quantum cannot be cloned. Nature, 299, 802–803.
Chapter 14
Is Quantum Mechanics a New Theory
of Probability?

Richard Healey

Abstract Some physicists but many philosophers believe that standard Hilbert
space quantum mechanics faces a serious measurement problem, whose solution
requires a new theory or at least a novel interpretation of standard quantum
mechanics. Itamar Pitowsky did not. Instead, he argued in a major paper (Pitowsky,
Quantum mechanics as a theory of probability. In: Demopoulos W, Pitowsky I (eds)
Physical theory and its interpretation. Springer, Dordrecht, pp 213–240, 2006) that
quantum mechanics offers a new theory of probability. In these and other respects
his views paralleled those of QBists (quantum Bayesians): but their views on the
objectivity of measurement outcomes diverged radically. Indeed, Itamar’s view of
quantum probability depended on his subtle treatment of the objectivity of outcomes
as events whose collective structure underlay this new theory of probability. I’ve
always been puzzled by the thesis that quantum mechanics requires a new theory of
probability, as distinct from new ways of calculating probabilities that play the same
role as other probabilities in physics and daily life. In this paper I will try to articulate
the sources of my puzzlement. I’d like to be able to think of this paper as a dialog
between Itamar and me on the nature and application of quantum probabilities.
Sadly, that cannot be: by taking his part in the dialog I will inevitably impose my
own distant, clouded perspective on his profound and carefully crafted thoughts.

Keywords Quantum probability · Dutch book argument · Measurement

problem · Quantum gamble · Gleason’s theorem · Observable outcomes ·
Pragmatism

R. Healey ()
Philosophy Department, University of Arizona, Tucson, AZ, USA
e-mail: rhealey@email.arizona.edu

© Springer Nature Switzerland AG 2020 317

M. Hemmo, O. Shenker (eds.), Quantum, Probability, Logic,
Jerusalem Studies in Philosophy and History of Science,
https://doi.org/10.1007/978-3-030-34316-3_14
318 R. Healey

14.1 Introduction

Itamar Pitowsky (2003, 2006) developed and defended the thesis that the Hilbert
space formalism of quantum mechanics is just a new kind of probability theory. In
developing this thesis he argued that all features of quantum probability, including
the Born probability rule, can be derived from rational probability assignments to
finite “quantum gambles”. In defense of his view he argued that all experimental
aspects of entanglement, including violations of the Bell inequalities, are explained
as natural outcomes of the resulting probability structure. He further maintained
that regarding the quantum state as a book-keeping device for quantum probabilities
dissolves the BIG measurement problem that afflicts those who believe the quantum
state is a real physical state whose evolution is always governed by a linear dynam-
ical law; and that a residual small measurement problem may then be resolved
by appropriate attention to the great practical difficulties attending measurement
of any observable on a macroscopic system incompatible with readily executable
measurements of macroscopic observables.
This is a bold view of quantum mechanics. It promises to transcend not only the
frustratingly inconclusive debate among so-called realist interpretations (including
Bohmian, Everettian, and “collapse” theories) but also the realist/instrumentalist
dichotomy itself. In this way it resembles the pragmatist approach I have been
developing myself (Healey 2012, 2017). Our views agree on more substantive
issues, including the functional roles of quantum states and probabilities and
the consequent dissolution of the measurement problem. But I remain puzzled
by Itamar’s central thesis. The best way to explain my puzzlement may be to
show how the points on which we agree mesh with a contrasting conception of
quantum probability. On this rival conception, quantum mechanics requires no new
probability theory, but only new ways of calculating probabilities with the same
formal features and the same role as other probabilities in physics and daily life.1
The rest of this contribution proceeds as follows. In the next section I say
what Pitowsky meant by a probability theory and explain how he took quantum
probability theory to differ formally from classical probability theory. The key
difference arises from the different event structures over which these are defined.
Whereas classical probabilities are defined over a σ -algebra of subsets of a set,
quantum probabilities are defined over the lattice L of closed subspaces of a Hilbert
space. Gleason’s theorem plays a major role here: for a Hilbert space of dimension
greater than 2, it excludes a classical truth-evaluation on L but at the same time
completely characterizes the class of possible quantum probability measures on L.
Pitowsky sought to motivate the formal difference between quantum and classical
probability theory by an extension of a Dutch book argument offered in support of
a coherence requirement on rational degrees of belief (in the tradition of Ramsey

1 Thatquantum mechanics involves a new way of calculating probabilities is a point made long
ago by Feynman, as Pitowsky himself noted. But Feynman (1951) also says that the concept of
probability is not thereby altered in quantum mechanics. Here I am in agreement with Feynman,
though the understanding he offers of that concept is disappointingly shallow.
14 Is Quantum Mechanics a New Theory of Probability? 319

(1926) and De Finetti (1937)) to a class of what he called “quantum gambles”

associated with possible measurements on a quantum system. Section 14.3 explains
the extension but questions how well it justifies the proposed formal modification.
A key issue here concerns the independence of the quantum probability of an event
from the context in which a non-maximal observable containing it is measured.
A quantum gamble may be settled only if there is a viable procedure for
determining which possible event assigned a quantum probability has occurred.
So quantum probabilities are not assigned to past events of which we have only
a partial record, or no record at all. This restricts “matters of fact” to include only
observable records. Section 14.4 notes how this is in tension with applications of
quantum theory in situations where we have no observable records. It goes on to
consider hypothetical situations to which quantum theory may be applied in which
records may be observable by some but not other observers before these records are
permanently erased so they can no longer be read by any observers.
In Sect. 14.5 I defend an alternative pragmatist view of probability, explaining
how it makes room for a notion of objective probability capable of rationally
constraining the credences of a rational agent. On this view probabilities in quantum
theory are defined over a family of Boolean algebras rather than a single non-
Boolean lattice. This is because Born probabilities are defined over possible
outcomes of a physical process represented in one, rather than some other, classical
event space. Models of decoherence may provide a guide as to what that process
is, even though they do not describe the dynamical evolution of a physical quantum
state. Any agent who accepts quantum theory should adjust his or her credences to
the Born probabilities given by the correct quantum state for one in that physical
situation. Insofar as an agent’s physical situation constrains what information is
available, differently situated agents are correct to assign different quantum states
and to adjust their credences accordingly. Each agent should update that quantum
state (and associated Born probabilities) on accessing newly available information.
This is one way a quantum state provides a book-keeping device for probabilities:
the other is provided by its linear evolution while the system remains undisturbed.

14.2 Quantum Measure Theory

Pitowsky maintained that quantum probability theory differs from classical proba-
bility theory formally as well as conceptually. Following Kolmogorov, probability
is usually characterized formally as a unit-normed, countably additive, measure Pr
on a σ -algebra  of subsets Ei (i ∈ N) of a set : that is

Pr :  −→ [0, 1] satisfies
1. Pr() = 1 (Probability measure)
 
2. Pr Ei = Pr Ei provided that ∀i = j (Ei ∩ Ej = ∅)
i i
320 R. Healey

Here  forms a lattice which is complemented and distributive, and hence a

Boolean algebra. Pr(Ei ) is the probability of the event Ei . Condition 2 is sometimes
weakened to finite additivity.
The closed subspaces {Si } of a Hilbert space H also form a lattice L(H ) whose
meet Si ∧ Sj is Si ∩ Sj , and whose join Si ∨ Sj is the smallest closed subspace
containing all elements of Si and Sj . H is the maximal element of L(H ) (denoted
by 1) and the null subspace ∅ is the minimal element, denoted by 0. Each element
S has a complement S ⊥ consisting of every vector orthogonal to every vector in S.
Indeed, S ⊥ is the unique orthocomplement of S. But this lattice is not distributive
and so L(H ) is not a Boolean algebra.
A (unit-normed) quantum measure μ on L(H ) is defined as follows:

μ : L(H ) −→ [0, 1] satisfies

1. μ(1) = 1 (Quantum measure)
 
2.μ( Si ) = Si provided that ∀i = j (Si ⊥ Sj )
i i

Condition 2 may be weakened to finite additivity.

The formal analogy between the conditions defining a probability measure and
those defining a quantum measure motivated Pitowsky’s proposal to interpret μ as a
new kind of quantum probability. Accordingly, he referred to the closed subspaces
of H as events, or possible events, or possible outcomes of experiments. He also
took himself to be following a tradition that takes L(H ) as the structure representing
the“elements of reality” in quantum theory. Regardless of that tradition, others have
also referred to quantum measures on non-Boolean lattices associated with quantum
mechanics as quantum probability functions or measures (Earman 2018; Ruetsche
and Earman 2012). The set of bounded self-adjoint operators on H forms a von
Neumann algebra B(H ): this includes as a subset the set P(B(H )) of projection
operators Ê on H . By virtue of the one-one correspondence between the set of
closed subspaces {S} of H and the operators {ÊS } that project onto them, P(B(H ))
also forms a lattice isomorphic to L(H ). This generalizes to von Neumann algebras
A other than P(B(H )) that figure in what Ruetsche (2011) calls extraordinary
quantum mechanics or QM∞ .
Gleason’s theorem (Gleason 1957) completely characterizes the class of quantum
measures on the lattice of closed subspaces of a Hilbert space of dimension greater
than 2.
Gleason’s Theorem: If H is a (separable) Hilbert space of dimension greater than 2 and μ
is a (unit-normed) quantum measure on L(H ), then there is a positive self-adjoint operator
Ŵ of trace 1 such that, for every element ÊS of P (B(H )), μ(S) = T r(Ŵ ÊS ).

This theorem has two important consequences. The first is that there are no
dispersion-free measures on L(H ) if dim(H ) 3, where a quantum measure μ
is dispersion-free if and only if its range is the set {0, 1}. This is important since
any truth-valuation on the set of all propositions stating that it is event S or
instead event S ⊥ that represents an element of reality would have to give rise to
14 Is Quantum Mechanics a New Theory of Probability? 321

a dispersion-free quantum measure on L(H ). This, of course, is why Gleason’s

theorem and similar results (e.g. Kochen and Specker (1967), Bell (1966)) are
considered important “no-hidden variable” results. While it may be understood as
the outcome of an experiment, an event S may not be (uniformly) understood as
simply the independently existing state of affairs that experiment serves to reveal.
The second important consequence of Gleason’s theorem is that every (unit-
normed) quantum measure on L(H ) (dim(H )  3) uniquely extends to the Born
probability measure associated with a quantum state, as represented by a vector in
or density operator on H . Indeed, this is true even if the quantum measure is merely
finitely additive. This consequence is not surprising, since the Born rule may be
stated in the form

Pr Ŵ (A ∈ ) = T r(Ŵ Ê()) (Born rule)

where the quantum state is represented by density operator Ŵ and Ê() is the
relevant projection operator from the spectral family defined by the unique self-
adjoint operator  corresponding to dynamical variable A.
Cognizant of the first consequence of Gleason’s theorem, Pitowsky did not
defend the radical thesis that quantum theory shows the world obeys a non-classical
logic. But he took the second consequence as “one of the strongest pieces of
evidence in support of the claim that the Hilbert space formalism is just a new kind
of probability theory” (Pitowsky 2006, p.222).

14.3 Quantum Gambles

Pitowsky developed his conception of quantum probability within the Bayesian

tradition pioneered by Ramsey (1926) and De Finetti (1937). This tradition locates
probabilities in an agent’s rationalized degrees of belief. A necessary, though
possibly insufficient, condition for such degrees of belief to be rational is that they
be (what has come to be called) coherent.
In this tradition, belief and its degrees are dispositions manifested not by self-
avowal but in actions. The condition of coherence is understood in terms of
the agent’s dispositions to accept a set of wagers at various odds offered by a
hypothetical bookie. A set of degrees of belief is coherent just in case it corresponds
to a set of such dispositions that is not guaranteed to result in a loss if collectively
manifested no matter what the outcome of all the bets in the set. Degrees of belief
that are not coherent are said to allow a Dutch book to be made against the agent
who has them. So-called synchronic Dutch book theorems are then taken to show
that any coherent set of degrees of belief in a set of propositions may be represented
by a (classical) probability measure over them.
Following Lewis (1980) an agent’s coherent set of degrees of belief are called
his or her credences. For a subjectivist like De Finetti, there is no further notion of
322 R. Healey

objective probability at which a rational agent’s credences should aim. Lewis (1980)
disagreed. Instead he located a distinct concept of objective probability he called
chance, of which all we know is that it provides a further rational constraint on an
agent’s credences through what he called the Principal Principle. I defer further
consideration of this principle until Sect. 14.5, since it plays no role in Pitowsky’s
own view of quantum probability theory.
According to Pitowsky (2006), a quantum gamble consists of four steps:
1. A single physical system is prepared by a method known to everybody.
2. A finite set M of incompatible measurements, each with a finite number of possible
outcomes, is announced by the bookie. The agent is asked to place bets on the possible
outcomes of each one of them.
3. One of the measurements in the set M is chosen by the bookie and the money placed on
all other measurements is promptly returned to the agent.
4. The chosen measurement is performed and the agent gains or loses in accordance with
his bet on that measurement.

Here M is identified with a set of Boolean algebras {B1 , B2 , . . . , Bk }, each

generated by the possible outcomes in L of the measurement to which it corre-
sponds. The elements of B = {S1 , S2 , . . . Sm } ∈ M will not all be one-dimensional
subspaces if M is not a maximal measurement. Pitowsky (2006, p.223) maintained
that “by acting according to the standards of rationality the gambler will assign
probabilities to the outcomes”. He took the gambler in question to recognize
identities in the logical structure consisting of the outcomes in L, and in particular
the cases in which the same outcome is shared by more than one experiment (i.e.
type of measurement in M.) But, crucially, this gambler was not assumed to know
quantum mechanics.
Pitowsky then argued (2006, p.227: Corollary 7) that any such quantum gambler
not meeting these standards of rationality must assign probability values to elements
of a finite sublattice 0 ⊂ L(H ) (dim(H ) ≥ 3) that cannot be extended to a quantum
measure on a finite  ⊃ 0 . He took this conclusion to establish the claim that
quantum theory is a new theory of probability.
Notice that this argument is not offered as a derivation of the Born rule insofar as
it does not mention the quantum state Ŵ used in stating that rule. It concludes only
that a rational agent’s credences should be consistent with the probabilities specified
by quantum mechanics through application of the Born rule to some quantum state.
Instead, this conclusion is presented as justification for the claim that quantum
probability is quantum measure theory: but is it warranted?
How do the standards of rationality constrain a quantum gambler’s assignments
of “probabilities” (i.e. degrees of belief) to outcomes in L? Assume that for
every B ∈ M such an agent assigns degree of belief cr(S|B) to outcome S of
a measurement corresponding to B. Pitowsky took the standards of rationality to
require the agent to assign degrees of belief in accordance with two rules (in my
notation):
RULE 1: For each measurement B ∈ M the function cr(•|B) is a probability distribution
on B.
RULE 2: If B1 , B2 ∈ M, and S ∈ B1 ∩ B2 then cr(S|B1 ) = cr(S|B2 ).
14 Is Quantum Mechanics a New Theory of Probability? 323

Rule 1 is imposed to insure the coherence of the agent’s degrees of belief

concerning the possible outcomes of a single measurement: If the agent’s degrees
of belief do not confirm to this rule they will dispose him or her to accept bets on
these outcomes guaranteed to lead to a sure loss. This is the standard Dutch book
argument as to why a rational agent’s degrees of belief must be representable as a
(finitely additive) probability measure.
The force of such Dutch book arguments has been a topic of extended debate
among philosophers of probability, and alternative arguments have been offered as
to why a rational agent’s degrees of belief should be representable as probabilities.
Briefly stated, here are three standard objections seeking to undermine the Dutch
book argument: A rational agent may simply refuse to bet: his betting behavior may
fail to reveal his degrees of belief insofar as it is a function of the rest of his cognitive
and affective state: agents have degrees of belief in propositions whose truth-value
cannot be determined because there is no corresponding settleable outcome. I will
not press either of the first two objections now. Pitowsky did anticipate this third
objection by requiring that there be a viable procedure for determining which
possible event assigned a quantum probability has occurred, and I will pursue this
issue in Sect. 14.4.
Rule 2 requires a quantum gambler’s credences to be non-contextual, in the sense
that he or she have the same degree of belief in the truth of propositions stating the
outcome of a measurement of a non-maximal observable no matter what type of
measurement led to that outcome. It is a remarkable fact about the Born rule of
quantum mechanics that the probabilities it yields are non-contextual in this sense.
But Pitowsky’s gambler cannot be assumed to know quantum mechanics. So why
should the quantum gambler’s credences be non-contextual?
Pitowsky (2006, p.216) first addresses this question in Sect. 2.1 as follows
(emphases in the original):
. . . the identity of events which is encoded by the structure also involves judgments of
probability in the sense that identical events always have the same probability. This is the
meaning of accepting a structure as an algebra of events in a probability space.

Here he takes the structure of events to be the lattice L(H ) of subspaces of

some Hilbert space H whose Boolean sublattices correspond to measurements
on a quantum system, some incompatible with others. Application of quantum
theory’s Born rule turns L(H ) into a quantum measure space when quantum state
Ŵ generates a subspace measure μ through μ(S) = T r(Ŵ P̂S ), where P̂S is the
(unique) projection operator onto subspace S ∈ L(H ). But so far we have been
offered no reason to regard L(H ) as a probability space.
The Dutch book argument Pitowsky took to require a rational quantum gambler’s
credences to conform to Rule 1 does not require that agent’s credences to conform
to Rule 2. To the extent that argument is successful, it establishes only that each
Boolean subalgebra B of L(H ) corresponding to a measurement in M may be
taken to define a σ -algebra of subsets of R representing possible outcomes of that
measurement, and in that sense B is a (subjective) probability space. Additional
324 R. Healey

argument is needed to justify the claim that the full lattice L(H ) is a probability
space.
The argument Pitowsky (2006, p.216) offers in Sect. 2.1 proceeds by analogy to
a classical application of probability theory to games of chance:
Consider two measurements A, B, which can be performed together; and suppose that A has
the possible outcomes a1 , a2 , . . . , ak , and B the possible outcomes b1 , b2 , . . . , br . Denote
by {A = ai } the event “the outcome of the measurement of A is ai ”, and similarly for
{B = bj }. Now consider the identity:


k
{B = bj } = ({B = bj } ∩ {A = ai }) (14.1)
i

This is the distributivity rule which holds in this case as it also holds in all classical cases.
This means, for instance, that if A represents the roll of a die with six possible outcomes
and B the flip of a coin with two possible outcomes, then Eq. (14.1) is trivial. Consequently
the probability of the left hand side of Eq. (14.1) equals the probability of the right hand
side, for every probability measure.

In a classical joint probability space the event {B = bj } occurs if and only if

some outcome {B = bj } ∩ {A = ai } occurs, and so the identity Eq. (14.1) holds
trivially. If instead the event {B = bj } is the outcome of a trial with a set of possible
outcomes {{B = bn } : n = 1, . . . , N }, then it is not trivial that the probability of the
left hand side of Eq. (14.1) equals the probability of the right hand side, since these
concern probabilities in different trials with different probability spaces. We would
be astonished if a fair coin always came up heads whenever a die was rolled at the
same time, but that is because we have empirical reasons to discount the influence
of dice rolls on the outcomes of simultaneous coin tosses.
Pitowski draws an analogy between Eq. (14.1) (as applied to compatible quantum
measurements or coin flips and dice rolls) and Eq. (14.2):


k 
l
({B = bj } ∩ {A = ai }) = {B = bj } = ({B = bj } ∩ {C = ci }) (14.2)
i i

In this equation, A, B, C, are quantum observables such that [A, B] = 0, and

[B, C] = 0 but [A, C] = 0, and c1 , c2 , . . . , cl are the possible outcomes of C. Since
A, C are incompatible observables they are not jointly measurable. Accordingly, the
Born rule of quantum mechanics does not assign joint probabilities to events such
as {A = am }, {C = cn }. Unlike the events {A = am }, {B = bn } in Eq. (14.1),
such events are not elements of any (classical) probability space acknowledged by
quantum mechanics. Indeed, as Fine (1982) showed, the joint Born probabilities for
some sets of pairwise compatible observables are not marginals of any (classical)
joint probability distribution. Equation (14.2) expresses a trivial identity in a lattice
L(H ) if ∪, ∩ are read as join and meet operations in the lattice. But they cannot
generally be read as set-theoretic union and intersection of elements of a σ -algebra
of subsets of a set of outcomes of a (classical) probability space.
14 Is Quantum Mechanics a New Theory of Probability? 325

In a way, this was Pitowsky’s point. He used the analogy only to motivate his
proposal that quantum probability is different from classical probability in just this
way. As he put it (Pitowsky 2006, pp.216–7)
I assume that the 0 of the algebra of subspaces represents impossibility (zero probability in
all circumstances) 1 represents certainty (probability one in all circumstances), and the iden-
tities such as Eqs. (14.1) and (14.2) represent identity of probability in all circumstances.
This is the sense in which the lattice of closed subspaces of the Hilbert space is taken as
an algebra of events. I take these judgments to be natural extensions of the classical case; a
posteriori, they are all justified empirically.

However, his argument for this proposal was supposedly based on rational-
ity requirements on the credences of a quantum gambler ignorant of quantum
mechanics, and here empirical considerations are out of place. If probability
just is rational credence, then a requirement of rationality cannot be based on
empirical considerations of the kind that are taken to warrant acceptance of quantum
mechanics and the non-contextuality of probabilities consequent on application of
the Born rule. Immediately after stating Rule 2, Pitowsky (2006) says that this
follows from an identity between events essentially equivalent to Eq. (14.2), and the
principle that identical events in a probability space have equal probabilities. But, as
we have seen, that principle applies automatically only when these events are part of
a single probability space, corresponding to a single trial or measurement context.
A lattice L(H ) is naturally understood as a quantum measure space, but only
its Boolean subspaces are naturally understood as probability spaces. Pitowsky’s
appeal to rationality requirements on a quantum gambler does not justify his thesis
that quantum theory is a new kind of probability theory.

14.4 Objective Knowledge of Quantum Events

In the language of probability theory, the term ‘event’ may be used to refer either to
a mathematical object (such as a set) or to a physical occurrence. When arguing that
the Hilbert space formalism of quantum mechanics is a new theory of probability,
Pitowsky took that theory to consist of an algebra of events, and the probability
measure defined on it. Events in this sense are mathematical objects—subspaces of a
Hilbert space. But to each such object he also associated a class of actual or possible
physical occurrences—outcomes of an experiment in which a quantum observable
is measured. A token physical event eS occurs just in case in a measurement M of
observable O the outcome is oi , an eigenvalue of Ô with eigenspace S. Here M is a
token physical procedure corresponding to the mathematical object BM , a Boolean
subalgebra of a lattice L(H ) of subspaces including S, where Ô is a self-adjoint
operator on H .
When a quantum gamble is defined over a set M of incompatible measurements,
each of these is characterized by the corresponding Boolean subalgebra of L(H ).
A token event eS settles a gamble if it is the outcome of an actual measurement
procedure M corresponding to the mathematical object BM ∈ M, where the bookie
326 R. Healey

chose to perform M rather than some other measurement in the gamble. Following
the subjectivist tradition pioneered by Ramsey, Pitowsky stressed that, for a quantum
gamble to reveal an agent’s degrees of belief, the outcome of whatever measurement
is chosen by the bookie must be settleable.
A proposition that describes a possible event in a probability space is of a rather special kind.
It is constrained by the requirement that there should be a viable procedure to determine
whether the event occurs, so that a gamble that involves it can be unambiguously decided.
This means that we exclude many propositions. For example, propositions that describe past
events of which we have only a partial record, or no record at all. (Pitowsky 2003, p.217)
Recall that in section 2.2 we restricted “matters of fact” to include only observable
records. Our notion of “fact” is analytically related to that of “event” in the sense that a
bet can be placed on x1 only if its occurrence, or failure to occur, can be unambiguously
recorded. (op. cit., p.231)

In the previous section I questioned the force of Pitowsky’s argument that a

rational agent’s degrees of belief in the outcomes of quantum gambles will be
representable as a quantum probability measure on L(H ). But even if that argument
were sound it would not extend to all uses of probability in applications of quantum
theory. The Born rule may be legitimately applied to propositions concerning
events whose outcomes are unknown or even unknowable. For such events there
is no viable procedure to determine what the outcome is, so there is no settleable
quantum gamble involving these events. An agent acting according to the standards
of rationality associated with a quantum gamble need not assign any probabilities
to the possible outcomes of such an event: rules 1 and 2 need not apply. Such a
rational agent may have degrees of belief concerning these outcomes that cannot
be represented as a quantum measure on a corresponding L(H ) even though each
outcome has a well-defined Born probability.
Quantum theory is often applied to occurrences in distant spacetime regions
from which no observable records are accessible. These include processes in the
center of stars (including the cores of neutron stars) and quantum field fluctuations
in the early universe. There is no restriction on the application of the Born rule
to calculate probabilities of outcomes associated with such processes. Of course
these cannot be understood as the outcomes of experiments since no experimenters
could have been present in those regions. But the Born rule is commonly applied
to yield probabilities of outcomes of processes in which no experiment is involved,
whether or not these are referred to as measurements. The notion of measurement
is notoriously obscure in quantum mechanics, as Bell forcefully pointed out in his
article “Against ‘Measurement”’. But even there he posed the rhetorical question
If the theory is to apply to anything but highly idealized laboratory operations, are we not
obliged to admit that more or less ‘measurement-like’ processes are going on more or less
all the time, more or less everywhere? (Bell 2004, p.216)

These days references to measurement are often replaced by talk of decoherence,

though it is now generally acknowledged that if there is a serious measurement
problem then appeals to decoherence will not solve it. While agreeing with Pitowsky
that there is no BIG measurement problem, in the next section I will indicate how
quantum models of decoherence may be used as guides to the legitimate application
14 Is Quantum Mechanics a New Theory of Probability? 327

of the Born rule in a pragmatist view of quantum mechanics. But in this view
a proposition to which the Born rule assigns a probability does not describe the
outcome of a measurement but the physical event of a magnitude taking on a value,
whether or not this is precipitated by measurement.
Even when a magnitude (a.k.a. observable) takes on a value recording the
outcome of a quantum measurement, that record may be subsequently lost, at
which point Pitowsky’s view excludes any event described by a proposition about
that outcome from a quantum probability space. While such scruples may not be
out of place in the context of a quantum gamble, only an indefensible form of
verificationism would prohibit retrospective application of the Born rule to that
event. So quantum mechanics permits application of the concept of probability in
circumstances in which it cannot be understood to be defined on a space of events
consistent with Pitowsky’s view.
One might object that no record of a measurement outcome is ever irretrievably
lost. But recent arguments (Healey 2018; Leegwater 2018) challenge such epistemic
optimism in the context of Gedankenexperimenten based on developments of
the famous Wigner’s friend scenario. These arguments threaten the objectivity
of measurement outcomes under the assumption that unitary, no–collapse single-
world quantum mechanics is applicable at all scales, even when applied to one
observer’s measurement on another observer’s lab (including any devices in that
lab recording the outcomes of prior quantum measurements). It is characteristic
of these arguments that an observation by one observer on the lab of another
completely erases all of the latter’s records of the outcome of his or her prior
quantum measurement. As Leegwater (2018, p.13) put it
such measurements in effect erase the previous measurement’s result; they must get rid of
all traces from which one can infer the outcome of the first measurement.

In this situation, the first observer’s measurement outcome was a real physical
occurrence, observable by and (then) known to that observer. But the second
observer’s measurement on the first observer’s lab erased all records of that
outcome, and neither of these observers, nor any other observer, can subsequently
verify what it was. Indeed, even the first observer’s memory of his own result has
been erased. This is clearly a situation in which Pitowsky would have excluded
the event corresponding to the first observer’s measurement outcome from any
probability space. For in no sense does a true proposition describing that outcome
state (what he called) a “matter of fact”.
I wonder what Pitowsky would have made of these recent arguments. Leegwater
(2018) presents his argument as a “no-go” result for unitary, no–collapse single-
outcome (relativistic) quantum mechanics, while Healey (2018) takes the third
argument he considers to challenge the objectivity of measurement outcomes.
Brukner (2018) formulates a similar argument in a paper entitled “A no-go theorem
for observer-independent facts”, in which he says
We conclude that Wigner, even as he has clear evidence for the occurrence of a definite
outcome in the friend’s laboratory, cannot assume any specific value for the outcome
to coexist together with the directly observed value of his outcome, given that all other
328 R. Healey

assumptions are respected. Moreover, there is no theoretical framework where one can
assign jointly the truth values to observational propositions of different observers (they
cannot build a single Boolean algebra) under these assumptions. A possible consequence of
the result is that there cannot be facts of the world per se, but only relative to an observer, in
agreement with Rovelli’s relative-state interpretation, quantum Bayesianism, as well as the
(neo)-Copenhagen interpretation.

Pitowsky founded his Bayesian approach to quantum probability on the notion

of a quantum gamble, from which he excluded propositions about measurement
outcomes that describe past events of which we have only a partial record, or no
record at all. This strongly suggests that he would be unwilling to countenance
applications of quantum theory to propositions about measurement outcomes stating
observer-dependent facts. His tolerance of agent-dependent probabilities did not
extend to acquiescence in non-objective facts about their subject matter. The paper
by Hemmo and Pitowsky (2007) criticized the way many-worlds interpretations of
quantum mechanics understood the use of probability in quantum mechanics. So
Pitowsky would almost certainly have rejected the option of evading the conclusion
of recent non-go arguments by countenancing multiple outcomes of a single
quantum measurement. I speculate that his response to these arguments would have
been to follow von Neumann by rejecting the assumption that unitary quantum
mechanics may be legitimately applied to the measurement process itself.

14.5 A Pragmatist View of Quantum Probability

As we saw in Sect. 14.3, Pitowsky’s Bayesian view of quantum probability relied

on his Rule 1, which requires an agent’s degrees of belief in the possible outcomes
of a quantum measurement characterized by a Boolean algebra B to be a (finitely
additive) probability measure on the associated algebra of sets. Here he followed in
the tradition of Ramsey and de Finetti, who first employed synchronic Dutch book
arguments in support of the probability laws as standards of synchronic coherence
for degrees of belief. But each of them also took an agent’s betting behavior
as a measure of his or her degrees of belief, giving operational significance to
the numerical probabilities by which these could be represented. Pitowsky quoted
Ramsey as follows:
“The old-established way of measuring a person’s belief ” by proposing a bet, and seeing
what are the lowest odds which he will accept, is “fundamentally sound”. (Pitowsky 2006,
p.223)

If degrees of belief are to be behaviorally defined in terms of betting behavior,

then it is essential that these bets be settleable—as Pitowsky insisted. But opera-
tional definitions of individual cognitive states like beliefs and desires or preferences
are now commonly regarded as inadequate. These are better understood within a
broadly functionalist approach to the mind in which the behavioral manifestations
of an individual belief are a function of many, if not all, an agent’s other cognitive
14 Is Quantum Mechanics a New Theory of Probability? 329

states. Even if there are such things as degrees of belief, these cannot be reliably
measured by an agent’s betting behavior of the kind that figures in an argument
intended to show that a (prudentially) rational agent’s degrees of belief will be
representable as probabilities.
A Dutch book argument may still be used to justify Bayesian coherence of a set
of degrees of belief as a normative condition on epistemic rationality, analogous to
the condition that a rational agent’s full beliefs be logically compatible. The view
that ideally rational degrees of belief must be representable as probabilities has been
called probabilism (Christensen 2004, p.107): such degrees of belief are known as
credences. The force of a Dutch book argument for probabilism is independent of
whether there are any bookies or whether bets in a book are settleable. Understood
this way, Rule 1 is justified as a condition of epistemic rationality on degrees
of belief in a quantum event, whether or not a gamble that involves it can be
unambiguously decided.
Probabilism places only minimal conditions on the degrees of belief of an indi-
vidual cognitive agent, just as logical compatibility places only minimal conditions
on his or her full beliefs. Pitowsky sought to justify additional conditions, sufficient
to establish the result that a rational agent’s credences in quantum events involving
a system be representable by a quantum measure on the lattice L(H ) of subspaces
of a Hilbert space H associated with that system (provided that dim(H )  3). This
was the key result he took to justify his claim that the Hilbert space formalism of
quantum mechanics is a new theory of probability.
Achieving this result depended on Rule 2: the further condition that a rational
agent’s credence in a quantum event represented by subspace S ∈ L(H ) be the
same, no matter whether it occurs as an outcome of measurement M1 (represented
by Boolean sub-lattice B1 ⊂ L(H )) or M2 (represented by B2 ). Pitowsky took
Rule 2 to follow from the identity between events, and the principle that identical
events in a probability space have equal probabilities. As embodied in Eq. (14.2),
Pitowsky took these judgments to be natural extensions of the classical case (as
embodied in (14.1)): and he said of these judgments “a posteriori, they are all
justified empirically”.
This justification for Rule 2 is quite different from the justification offered
for Rule 1, which was justified not empirically but as a normative requirement
on epistemic rationality. As such, Rule 1 showed why an epistemically rational
agent should adopt degrees of belief representable as probabilities over a classical
probability space of events associated with each Boolean subalgebra B of L(H ):
in that sense, it justified considering B as a probability space. It is because no
analogous normative requirement justifies taking L(H ) itself to be a probability
space that Pitowsky appealed instead to empirical considerations. But while such
empirical considerations may justify the introduction of a quantum measure μ
over L(H ) as a convenient device for generating a posteriori justified objective
probability measures over Boolean subalgebras of L(H ), this does not show that
μ is itself a probability, or L(H ) a probability space. Only by appeal to some
additional normative principle of epistemic rationality could a Bayesian in the
tradition of Ramsey and de Finetti attempt to show that.
330 R. Healey

The bearing of empirical considerations on credences has been a controversial

issue among Bayesians. De Finetti maintained that there is nothing more to
probability than each agent’s actual credences. Ramsey allowed that probability in
physics may require more, and in this he has been followed by the physicist Jaynes
(2003) and other objective Bayesians. Contemporary QBists portray the Born rule as
an empirically motivated additional normative constraint on credences (Fuchs and
Schack 2013), while most physicists still follow Feynman in seeking an objective
correlate of probability in stable frequencies of outcomes in repeated experimental
trials.
Lewis believed that objective probability required the kind of indeterminism
generally thought to be manifested by radioactive decay. He took orthodox quantum
mechanics to involve such indeterminism, with objective probabilities supplied by
the Born rule serving as paradigm instances of what he called chance. Lewis (1980)
formulated the Principal Principle he took to state all we know about chance by
linking this to credence. He later proposed a modification to square it with his
Humean metaphysics. But Ismael (2008) argued that the modification was a mistake,
and essentially restated his original principle. In Lewis’s (1994, pp.227–8) words,
If a rational believer knew that a chance of [an event e] was 50%, then almost no matter
what he might or might not know as well, he would believe to degree 50% that [e] was
going to occur. Almost no matter, because if he had reliable news from the future about
whether e would occur, then of course that news would legitimately affect his credence.

There is now a large literature on how a Principal Principle should be stated, and
how, if at all, it may be justified. This includes Ismael’s helpful formulation as an
implicit definition of a notion of chance:
The [modified Lewisian] chance of A at p, conditional on any information H p about the
contents of p’s past light cone satisfies: Crp (A/Hp ) =df Chp (A).

Lewis was right to add to credence a second, more objective, concept of

probability: but he was wrong to restrict its application to indeterministic contexts.
The physical situation of a localized agent frequently imposes limitations on access
to relevant information about an event about whose occurrence he or she wishes to
form reasonable credences. This occurs, for example, in so-called games of chance
and in classical statistical mechanics as well as in the quantum domain. In such
situations general principles or a physical theory may provide a reliable way of
“packaging” the accessible information in a way that permits generally reliable
inferences and appropriate actions. That is how I understand the role of the Born
rule in quantum mechanics. To accept quantum mechanics is to grant the Born rule
objective authority over one’s credences in quantum events and thereby regard it as
epistemic expert in this context.
The Born rule associates a set of general probabilities to events of certain types
involving a kind of physical system assigned a quantum state. The physical situation
of an actual or merely hypothetical agent gives that agent access to information
about the surrounding circumstances that may be sufficient to assign a specific
quantum state to one or more individual systems. Differently situated agents should
sometimes correctly assign different quantum states because different information
14 Is Quantum Mechanics a New Theory of Probability? 331

is accessible to each. This may occur just because the agents do not share a single
spacetime location, in conformity to Ismael’s modified Lewisian chance: Born
probabilities supply many examples of such objective probabilities. After a specific
quantum state is assigned to an individual system, the Born rule yields a probability
measure over events of certain types involving it. By instantiating the general Born
rule, the agent may then derive an objective chance distribution over possible events.
This is a classical probability distribution over an event space with the structure
of a Boolean algebra B, not a quantum measure over a non-Boolean lattice: in
a legitimate application of the Born rule no probability is assigned to events
that are not elements of B. Events assigned a probability in this way are given
canonical descriptions of the form Qs ∈ , where Q is a dynamical variable
(observable), s is a quantum system, and  is a Borel set of real numbers: I call
Qs ∈  a magnitude claim. Some of these events are appropriately redescribed as
measurement outcomes: for others, this redescription is less appropriate.
Probability theory had its origins in a dispute concerning dice throws, so it may
be helpful to draw an analogy with applications of probability theory to throws
of a die. The faces of a normal die are marked in such a way that the point total
of opposite faces sums to 7, and the faces marked 4, 6 meet on one edge of the
die. Consider an evenly weighted die that includes a small amount of magnetic
material carefully distributed throughout its bulk. This die may be thrown onto a
flat, level surface in one or other of two different kinds of experiments: a magnetic
field may be switched on or off underneath the surface as the die is thrown. In an
experiment with the magnetic field off, a throw of the die is fair, so there is an equal
chance of 1/6 that each face will land uppermost. (Notice that, though natural, this
use of the term ‘chance’ does not conform to Lewis’s Principal Principle—even
in Ismael’s modified formulation—since it does not presuppose that dice throws
are indeterministic processes.) In an experiment with the magnetic field on, the die
is biassed so that some faces are more likely to land uppermost than others. But
because of the careful placement of the magnetic material within the die, the chance
is still 1/3 that a face marked 4 or 6 lands uppermost.
In each kind of experiment, a probabilistic model of dice-throwing may be tested
against frequency data from repeated dice throws of that kind. This model includes
general probability statements of the form PX (E) = p specifying the probability
of an event of type E describing a possible outcome of a throw of the die in an
experiment of type X. Consider the situation of an actual or hypothetical agent
immediately prior to an individual throw of the die. The information accessible
to this agent does not include the outcome of that throw: nor could it include
all potentially relevant microscopic details of the initial and boundary conditions
present in that throw, even if it were a deterministic process. But it does include
knowledge of whether the magnetic field is on or off: failure to obtain this relevant,
accessible information would be an act of epistemic irresponsibility.
Having accepted a probabilistic model on the evidence provided by relative
frequencies of different outcome types in repeated throws of both kinds, this actual
or hypothetical agent has reason to instantiate the general probability statement
of the form PX (E) = p for each possible event e of type E in experiment x of
332 R. Healey

kind X. The result is the chance of e in x—that to which an epistemically rational

agent should match his or her credence concerning event e in experiment x. What
that chance is may depend on the experiment x. If e is an event of face 1 landing
uppermost, then the chance of e may be less if the magnetic field is on in x than it
would have been if the field had been off. But if e is the event of a prime-numbered
face landing uppermost, then the chance of e is 2/3 both in experiment x (magnetic
field off) and in experiment x  (magnetic field on).
Here is the analogy between this dice-throwing example and quantum probabili-
ties. In both cases there are general probability rules that may be instantiated to give
chances to which an (actual or hypothetical) rational agent who accepts those rules
should match his or her credences in circumstances of the kind specified by the rule.
In both cases these circumstances may be described as those of an experiment, but
in neither case is this description essential—no experimenter need be present if the
circumstances occur naturally.
In both cases a possible event of a certain type may be assigned the same chance
in different circumstances, describable as experiments of different kinds. Both cases
feature non-contextual probability assignments to events of the same type. There is
no temptation to combine the classical probability spaces corresponding to different
kinds of experiment in the dice-throwing example into a single non-classical event
space. The rich formal structures to which Pitowsky appealed in the quantum case
have no analog in the dice-throwing example. But I remain unpersuaded that the
absence of any corresponding structures in the dice-throwing example undermines
the force of the analogy.
It is the physical circumstances in which a system finds itself that determine
which applications of the Born rule are legitimate (if any), and which are not.
Quantum mechanics itself does not specify these circumstances, but physicists have
developed reliable practical knowledge of when they obtain. Quantum models of
decoherence can provide useful guidance in judging whether and which application
of the Born rule is legitimate, by selecting an appropriate Boolean algebra B of
events corresponding to a so-called “pointer basis” of subspaces of the Hilbert space
Hs of s. Here s will typically differ from the target system t to which a quantum state
has been assigned in order to apply the Born rule.
Models of decoherence are neither sufficient nor necessary to solve any BIG
measurement problem: there is no such problem. That problem would arise only on
the mistaken assumption that a quantum state specifies all the physical properties
of the system to which it is assigned. But I agree with Pitowsky that a quantum
state does not do this: instead, it acts as a book-keeping device for a rational agent’s
credences by requiring them to conform to the associated Born rule probabilities.
Quantum mechanics cannot explain why measurements have definite outcomes,
since application of its Born rule presupposes that they do. But since a quantum
state does not specify its physical condition, there is no tension between a system’s
being assigned a superposed state and the truth of a magnitude claim that it has a
particular eigenvalue of the corresponding operator.
The Born rule is legitimately applied only to significant magnitude claims. The
significance of a magnitude claim is not certifiable in terms of truth-conditions
14 Is Quantum Mechanics a New Theory of Probability? 333

in some kind of “quantum semantics”. In my pragmatist view, the significance

of any claim arises from its place in an inferential web of statements, ultimately
linked to perception and action through practical rather than theoretical inference.
This inferentialist “metasemantics” supports an analog rather than digital view of
the content of each magnitude claim as coming in degrees rather than making the
claim simply meaningful or meaningless. Though extraordinarily rapid, complete,
and practically irreversible, decoherence is also a matter of degree in applicable
quantum models. It follows that there can be no sharp line dividing meaningful
from meaningless magnitude claims, or legitimate from illegitimate applications of
the Born rule.
This sheds light on the import of the arguments discussed in the previous section
purporting to show that the universal applicability of unitary quantum mechanics is
incompatible with the assumption that quantum measurements always have unique,
objective outcomes. In my pragmatist view, the outcome of a quantum measurement
can be stated in a magnitude claim Qs ∈  about a system s that may be thought
of as a measuring apparatus. The claim Qs ∈  is true at a time if and only if (the
value of) Qs is then an element of . But Qs ∈  derives its content not through
this (trivial) truth-condition, but through the reliable inferences into which the claim
then enters.
In ordinary circumstances we assume that observation of an experimenter’s lab-
oratory records is a reliable way for anyone to determine the outcome of a quantum
measurement in that laboratory. Indeed, that is a common understanding of what it
is for the measurement to have a unique, objective outcome. But this assumption
breaks down in the extreme circumstances of Gedankenexperimenten like those that
figure in arguments considered in Healey (2018) and Leegwater (2018). In such
circumstances an observer in an initially physically isolated laboratory will correctly
report the outcome of his experiment even though another observer may report a
different outcome after subsequently entering that laboratory and making her own
observations of its records. Moreover these observers will not then register any
disagreement, since all of the first observer’s initial records (including memories)
will have been erased by the time of the second observation.
In such cases, processes modeled by decoherence confined to the first observer’s
laboratory rendered reliable a host of inferences based on magnitude claims stating
records of his quantum measurement. This endowed these claims with a high degree
of significance, so his observations warranted him in taking them truly to state the
unique, physical outcome of his measurement. But the extraordinary circumstances
of the Gedankenexperiment in fact restrict the domain of reliability to the context
internal to the laboratory within which these processes were confined. There is a
wider context that includes subsequent physical processes coupling that laboratory
to a second observer and her laboratory. In that wider context, inferences based
on the magnitude claims stating records of his quantum measurement cease to be
reliable, thereby curtailing the significance of these claims, which were nevertheless
true in the context in which he made them.
In my pragmatist view, quantum measurements would have unique, physi-
cal outcomes even in extraordinary circumstances like those described in the
334 R. Healey

Gedankenexperimenten that figure in the arguments considered in Healey (2018)

and Leegwater (2018). The outcomes could be described by true magnitude claims
about individual experimenters’ physical records of them. Some such claims made
by different experimenters may seem inconsistent. But the inconsistency is merely
apparent. A correct understanding of these claims requires that their content be
relativized to the physical context to which they pertain. So contextualized, each
experimenter in one of these Gedankenexperimenten can allow the objective truth of
the others’ reports of each of their unique measurement outcomes while consistently
stating the outcome of his or her own measurement.
This implicit limitation on the content of observation reports of the outcomes
of quantum measurements may usually be neglected. Quantum decoherence is so
pervasive that we will never be able to realize the extraordinary circumstances
required by the Gedankenexperimenten that figure in the arguments considered in
Healey (2018) and Leegwater (2018): even a powerful quantum computer would not
constitute an agent capable of performing and reporting the outcome of a quantum
measurement in a physically isolated laboratory. Because we all inevitably share
a single decoherence context, true reports of the unique, physical outcomes of
quantum measurements provide the objective data that warrant acceptance of the
theory we use successfully to predict their objective probabilities.

14.6 Conclusion

Quantum mechanics is not a new theory of probability. On the contrary, it constitutes

perhaps our most successful deployment of classical probability theory in physics.
It is not only the mathematics of probability that are classical here: the concept
itself functions in basically the same way in quantum mechanics that it always has.
This function is as a source of expert, “pre-packaged” advice to an actual or merely
hypothetical situated agent on how strongly to believe statements whose truth-value
that agent is not in a position to determine from the accessible information.
Quantum probabilities may be represented by means of quantum measures on
non-Boolean lattices, in which case Gleason’s theorem offers an elegant character-
ization of the range of quantum probabilities yielded by applications of the Born
rule. But a quantum measure is not a probability measure, and the lattice of closed
subspaces of a Hilbert space is not a probability space, despite the non-contextuality
of quantum probabilities.
Dutch book arguments for synchronic coherence support an epistemic norm that
an agent’s degrees of belief in a set of propositions should be representable by a
probability measure. But there is a more objective notion of probability than that
of an individual agent’s actual credences. An agent’s credences become subject
to additional norms through acceptance of general probabilities: these importantly
include the Born probabilities prescribed by a legitimate application of the Born
rule. Coherence is an epistemic norm even for beliefs concerning unsettleable
events, including the outcomes of quantum ‘measurements’ that no-one knows, and
14 Is Quantum Mechanics a New Theory of Probability? 335

those not everyone can know—but even these outcomes are as objective as science
needs them to be.
In all these ways I have come to disagree with Itamar’s view of probability in
quantum mechanics. I only wish he could now reply to show me why I am wrong:
we could all learn a lot from the ensuing debate. Let me finish by reiterating two
important points of agreement. There is no BIG quantum measurement problem:
there is only the small problem of physically modeling actual measurements within
quantum theory and showing why some are much easier than others. A quantum
state is not an element of physical reality (|ψ is not a beable); it is a book-keeping
device for updating an agent’s credences as time passes or in the light of new
information (even if there is no actual agent or bookie to keep the books)!

References

Bell, J. S. (1966). On the problem of hidden variables in quantum mechanics. Reviews of Modern
Physics, 38, 447–52.
Bell, J. S. (2004). Speakable and unspeakable in quantum mechanics (2nd ed.). Cambridge:
Cambridge University Press.
Brukner, Č. (2018). A no-go theorem for observer-independent facts. Entropy, 20, 350.
Christensen, D. (2004). Putting logic in its place. Oxford: Oxford University Press.
De Finetti, B. (1937). La prévision: ses lois logiques, ses sources subjectives. Annales de l’Institute
Henri Poincaré, 7, 1–68.
Earman, J. (2018). The relation between credence and chance: Lewis’ “Principal Principle” is a
theorem of quantum probability theory. http://philsci-archive.pitt.edu/14822/. Accessed 6 May
2019.
Feynman, R. P. (1951). The concept of probability in quantum mechanics. In Second Berkeley Sym-
posium on Mathematical Statistics and Probability, 1950 (pp. 533–541). Berkeley: University
of California Press.
Fine, A. (1982). Joint distributions, quantum correlations, and commuting observables. Journal of
Mathematical Physics, 23, 1306–1310.
Fuchs, C., & Schack, R. (2013). Quantum-Bayesian coherence. Reviews of Modern Physics, 85,
1693–1715.
Gleason, A. M. (1957). Measures on the closed subspaces of a Hilbert space. Journal of
Mathematics and Mechanics, 6, 885–893.
Healey, R. A. (2012). Quantum theory: a pragmatist approach. British Journal for the Philosophy
of Science, 63, 729–71.
Healey, R. A. (2017). The quantum revolution in philosophy. Oxford: Oxford University Press.
Healey, R. A. (2018). Quantum theory and the limits of objectivity. Foundations of Physics, 48,
1568–89.
Hemmo, M., & Pitowsky, I. (2007). Quantum probability and many worlds. Studies in History and
Philosophy of Modern Physics, 38, 333–350.
Ismael, J. T. (2008). Raid! Dissolving the big, bad bug. Nous, 42, 292–307.
Jaynes, E. T. (2003). In G. Larry Bretthorst (Ed.), Probability theory: The logic of science.
Cambridge: Cambridge University Press.
Kochen, S., & Specker, E. (1967). The problem of hidden variables in quantum mechanics. Journal
of Mathematics and Mechanics, 17, 59–87.
Leegwater, G. (2018). When Greenberger, Horne and Zeilinger meet Wigner’s friend.
arXiv:1811.02442v2 [quant-ph] 7 Nov 2018. Accessed 6 May 2019.
336 R. Healey

Lewis, D. K. (1980). A subjectivist’s guide to objective chance. In R. C. Jeffrey (Ed.), Studies in

inductive logic and probability (Vol. II, pp. 263–293). University of Berkeley: California Press.
Lewis, D. K. (1994). Humean supervenience debugged. Mind, 103, 473–90.
Pitowsky, I. (2003). Betting on the outcomes of measurements: A Bayesian theory of quantum
probability. Studies in History and Philosophy of Modern Physics, 34, 395–414.
Pitowsky, I. (2006). Quantum mechanics as a theory of probability. In W. Demopoulos & I.
Pitowsky (Eds.), Physical theory and its interpretation (pp. 213–240). Dordrecht: Springer.
Ramsey, F. P. (1926). Truth and probability. Reprinted In D. H. Mellor (Ed.), F. P. Ramsey:
Philosophical Papers. Cambridge: Cambridge University Press. 1990.
Ruetsche, L. (2011). Interpreting quantum theories. Oxford: Oxford University Press.
Ruetsche, L., & Earman, J. (2012). Infinitely challenging: Pitowsky’s subjective interpretation
and the physics of infinite systems. In Y. Ben-Menahem & M. Hemmo (Eds.), Probability
in physics. Berlin: Springer.
Chapter 15
Quantum Mechanics as a Theory
of Probability

Meir Hemmo and Orly Shenker

Abstract We examine two quite different threads in Pitowsky’s approach to the

measurement problem that are sometimes associated with his writings. One thread
is an attempt to understand quantum mechanics as a probability theory of physical
reality. This thread appears in almost all of Pitowsky’s papers (see for example
2003, 2007). We focus here on the ideas he developed jointly with Jeffrey Bub
in their paper ‘Two Dogmas About Quantum Mechanics’ (2010) (See also: Bub
(1977, 2007, 2016, 2020); Pitowsky (2003, 2007)). In this paper they propose an
interpretation in which the quantum probabilities are objective chances determined
by the physics of a genuinely indeterministic universe. The other thread is sometimes
associated with Pitowsky’s earlier writings on quantum mechanics as a Bayesian
theory of quantum probability (Pitowsky 2003) in which the quantum state seems to
be a credence function tracking the experience of agents betting on the outcomes of
measurements. An extreme form of this thread is the so-called Bayesian approach
to quantum mechanics. We argue that in both threads the measurement problem
is solved by implicitly adding structure to Hilbert space. In the Bub-Pitowsky
approach we show that the claim that decoherence gives rise to an effective Boolean
probability space requires adding structure to Hilbert space. With respect to the
Bayesian approach to quantum mechanics, we show that it too requires adding
structure to Hilbert space, and (moreover) it leads to an extreme form of idealism.

Keywords Bub-Pitowsky’s approach · Bayesian approach · measurement

problem · preferred basis problem

M. Hemmo ()
Department of Philosophy, University of Haifa, Haifa, Israel
e-mail: meir@research.haifa.ac.il
O. Shenker
Edelstein Center for History and Philosophy of Science, The Hebrew University of Jerusalem,
Jerusalem, Israel
e-mail: Orly.shenker@mail.huji.ac.il

© Springer Nature Switzerland AG 2020 337

M. Hemmo, O. Shenker (eds.), Quantum, Probability, Logic,
Jerusalem Studies in Philosophy and History of Science,
https://doi.org/10.1007/978-3-030-34316-3_15
338 M. Hemmo and O. Shenker

15.1 Introduction

The measurement problem in quantum mechanics is the problem of explaining how

individual measurement outcomes are brought about, given the Hilbert space struc-
ture of the quantum state space and the unitary dynamics (the Schrödinger equation
in the non-relativistic case). Influential interpretations of quantum mechanics such
as Bohm’s (1952) ‘hidden variables’ theory and the dynamical collapse theory
by Ghirardi, Rimini and Weber (1986; see also Bell 1987a) propose solutions to
this problem by adding assumptions about the structure of physical reality and
the dynamical equations of motion over and above the Hilbert space structure,
respectively, Bohmian trajectories and GRW collapses (or flashes; see Tumulka
2006). By contrast, Pitowsky’s approach to quantum mechanics as a (non-classical)
theory of probability aims at solving (or dissolving) the measurement problem,
without adding anything to Hilbert space.1
In this paper we examine two quite different threads in Pitowsky’s approach to
the measurement problem that are sometimes associated with his writings. One
thread is an attempt to understand quantum mechanics as a theory of probability
which is about physical reality. This thread appears in almost all of Pitowsky’s
papers (see for example 2003, 2007). We shall focus here on the ideas he developed
jointly with Jeffrey Bub in their paper ‘Two Dogmas About Quantum Mechanics’
(2010).2 In this paper they propose an interpretation in which the quantum probabil-
ities are objective chances determined by the physics of a genuinely indeterministic
universe. The other thread is sometimes associated with Pitowsky’s earlier writings
on quantum mechanics as a Bayesian theory of quantum probability (Pitowsky
2003) in which the quantum state seems to be a credence function tracking the
experience of agents betting on the outcomes of measurements. An extreme form of
this thread is the so-called Bayesian approach to quantum mechanics, which as we
read it addresses only the experience of agents.3 We argue that in both approaches
the measurement problem is solved by implicitly adding structure to Hilbert space.
In the Bub-Pitowsky approach we show that the claim that decoherence gives rise
to an effective Boolean probability space requires adding structure to Hilbert space.
With respect to the Bayesian approach to quantum mechanics, we show that it too
requires adding structure to Hilbert space, and (moreover) it leads to an extreme
form of idealism.
The paper is structured as follows. In Sect. 2 we briefly describe the Bub-
Pitowsky approach. In Sect. 3 we describe the role of decoherence in their approach
and how it gives rise to an effectively classical probability space of the events
associated with our familiar experience. But we argue that this falls short of
providing what they call a ‘consistency proof’ that the quantum (unitary) dynamics

1 Everett’s (1957) interpretation of quantum mechanics has a similar attempt in this respect.
2 See also: Bub (1977, 2007, 2016, 2020); Pitowsky (2003, 2007).
3 For versions of the Bayesian approach to quantum mechanics, see e.g., Caves et al. (2002a, b,

2007); Fuchs and Schack, (2013); Fuchs et al. (2014); Fuchs and Stacey (2019).
15 Quantum Mechanics as a Theory of Probability 339

is consistent with our experience in a way that is analogous to special relativity.

Finally, in Sect. 4 we criticise the Bayesian approach to quantum mechanics.

15.2 The Bub-Pitowsky Approach

According to Bub and Pitowsky (2010), quantum mechanics is primarily a theory of

non-classical probability describing a genuinely indeterministic universe in which
“probabilities (objective chances) are uniquely given from the start” by the geometry
of Hilbert space (Bub and Pitowsky 2010, p. 444; see also von Neumann 2001, p.
245). The event space in this theory of probability is the projective geometry of
closed subspaces of a Hilbert space, which form an infinite family of intertwined
Boolean algebras, but cannot be embedded into a single Boolean algebra. This space
imposes built-in objective constrains on the possible configurations of individual
events (see Pitowsky 2003, 2007; Bub and Pitowsky 2010). The unitary evolution
of the quantum state evolves the entire structure of correlations between events in
the Hilbert space, but it does not describe the transition from one configuration of
individual events to another (a special case of which is measurement). By Gleason’s
theorem, the Hilbert space structure determines unique probabilities of individual
events (e.g., measurement outcomes) which are encoded in the quantum state. They
take this to imply that the quantum state is a credence function (Bub and Pitowsky
2010, p. 440):
[T]he quantum state is a credence function, a bookkeeping device for keeping track of
probabilities – the universe’s objective chances . . . (Bub and Pitowsky 2010, p. 441; our
emphasis)

In their approach, von Neumann’s projection postulate turns out to stand for
a rule of updating probabilities by conditionalisation on a measurement outcome,
that is the von Neumann-Lüders rule. The fact that the updating is non-unitary
expresses in their view the ‘no cloning’ principle in quantum mechanics, according
to which there is no universal cloning machine capable of copying the outputs of an
arbitrary information source. This principle entails, via Bub and Pitowsky’s (2010,
p. 455) “information loss theorem”, that a measurement necessarily results in the
loss of information (e.g., the phase in the pre-measurement state). Since the ‘no
cloning’ principle is implied by the Hilbert space structure, they argue, the loss of
information in measurement is independent of how the measurement process might
be implemented dynamically. By this they mean that if no structure is added to
Hilbert space, any dynamical quantum theory of matter and fields that will attempt
to account for the occurrence of individual events must satisfy this principle (and
other constraints imposed by the quantum event space, such as ‘no signaling’4 ),
regardless of its details.

4 The ‘no signaling’ principle asserts, roughly, that the marginal probabilities of events associated
with a quantum system are independent of the particular set of mutually exclusive and collectively
exhaustive events associated with any other system.
340 M. Hemmo and O. Shenker

This idea, that quantum mechanics is a theory of probability in which the event
space is Hilbert space, leads them to reject what they identify as two ‘dogmas’ of
quantum mechanics: that measurement is a process that, as a matter of principle,
should always be open to a complete dynamical analysis; and that a quantum pure
state is analogous to the classical state given by a point in phase space, that is
the quantum state is a description (or representation) of physical reality. On the
view they propose, the quantum state is a credence function, which keeps track
of the objective chances in the universe, it is not the ‘truthmaker’ of propositions
about the occurrence or non-occurrence of individual events; and measurement is
primitive in the sense that the no-cloning principle follows from the non-Boolean
structure of the event space, and if one does not add structure to Hilbert space (like
Bohmian trajectories, GRW collapses), then there is no further deeper story to be
told about how measurement outcomes are brought about dynamically. In this sense
the measurement problem, the ‘big’ problem as they call it, is a pseudo problem.
But why is there no deeper story to be told? After all, Bub and Pitowsky
themselves realise that the quantum event space imposes constrains on the ‘objec-
tive chances’ of all possible configurations of individual events, and the unitary
dynamics evolves this entire structure: it does not describe the evolution from one
configuration of the universe to another. But if the theory does not describe the
transition from one configuration to the next, what are the ‘objective chances’
chances of? And in what sense is this quantum theory of probability a physical
theory of an indeterministic universe? In fact, in what sense is the universe
indeterministic at all, given the determinism of the unitary dynamics (which is all
there is)?
To these questions, Bub and Pitowsky reply by an analogy from special relativity:
If we take special relativity as a template for the analysis of quantum conditionalization
and the associated measurement problem, the information-theoretic view of quantum
probabilities as ‘uniquely given from the start’ by the structure of Hilbert space as a
kinematic framework for an indeterministic physics is the proposal to interpret Hilbert
space as a constructive theory of information-theoretic structure or probabilistic structure,
part of the kinematics of a full constructive theory of the constitution of matter, where
the corresponding principle theory includes information-theoretic constraints such as ‘no
signaling’ and ‘no cloning.’ Lorentz contraction is a physically real phenomenon explained
relativistically as a kinematic effect of motion in a non-Newtonian spacetime structure.
Analogously, the change arising in quantum conditionalization that involves a real loss of
information is explained quantum mechanically as a kinematic effect of any process of
gaining information of the relevant sort in the non-Boolean probability structure of Hilbert
space (irrespective of the dynamical processes involved in the measurement process).
Given ‘no cloning’ as a fundamental principle, there can be no deeper explanation for the
information loss on conditionalization than that provided by the structure of Hilbert space
as a probability theory or information theory. The definite occurrence of a particular event
is constrained by the kinematic probabilistic correlations encoded in the structure of Hilbert
space, and only by these correlations – it is otherwise ‘free’. (Bub and Pitowsky, p. 448)

On their account, the Hilbert space structure is analogous to the Minkwoski

spacetime5 ; and the quantum conditionalisation is analogous to the physically real

5 Or Einstein’s principle of relativity and the light postulate.

15 Quantum Mechanics as a Theory of Probability 341

Lorentz contraction (the contraction is different in different Lorentz frames, but

all the frames are equally real). Since a conditionalisation is expressed in terms
of some specific basis states in the Hilbert space, all choices of basis are (prima
facie) equally real, presumably in an analogous sense to the equal status of all
Lorentz frames in special relativity. It seems to us that in the quantum case, the deep
explanation of the conditionalisation in measurement – which, by the Bub-Pitowsky
analogy, is also physically real – is the non-Boolean probabilistic structure of the
Hilbert space, which includes the ‘no cloning’ principle. By the analogy, dynamical
analyses of the information loss in measurement are to be seen as attempts to solve
a consistency problem: that is, to find a dynamical analysis of the conditionalisation
upon measurement that is consistent with the Hilbert space structure. We shall come
back to this point later.
Bub and Pitowsky argue that dynamical analyses of the breaking of the thread
in Bell’s (1987b) spaceships scenario (an example which Bub and Pitowsky
2010 consider) in terms of Lorentz covariant forces is a consistency proof that
a relativistic dynamics compatible with the geometrical structure of Minkowski
spacetime is possible.6 The breaking of the thread in this scenario is a real fact
about which all the Lorentz frames agree (although the fact is explained differently
in different frames). Bub and Pitowsky’s claim is that the deep explanation for the
breaking of the thread is the Lorentz contraction, which is a feature of the geometry
of Minkowski spacetime that is independent of the material constitution of the
thread and the kind of interactions involved.
What is the quantum analogue of the breaking of the thread? There are three
candidates that come to mind (perhaps there are more): (i) The fact that a
measurement has a definite outcome in our experience. Bub and Pitowsky do not
attempt to explain this fact, so we set this issue aside. (ii) The fact that the structure
of the probability space of the events in our experience seems to be classical
(Boolean). Bub and Pitowsky set this fact (ii) as their desideratum. We shall criticise
their approach on this particular point in the next section. Another candidate is: (iii)
The objective basis-independent fact that the Hamiltonian of the universe includes
decoherence interactions.
In the literature, there are quantum mechanical theories that provide dynamical
analyses of measurement, such as Bohm’s (1952) theory and the GRW (1986) the-
ory. But these theories add structure to Hilbert space, and in virtue of this structure
they solve the measurement problem. For this reason Bub and Pitowsky (2010, p.
454) take Bohm’s theory and the GRW theory as “‘Lorentzian’ interpretations of
quantum mechanics”, which assume the Newtonian spacetime structure and add to
it the ether as an additional structure for the propagation of electromagnetic effects.’
Because they add structure, these ‘Lorentzian’ theories are not solutions to the
consistency problem (of finding a dynamical analysis that is compatible with the
Hilbert space structure (they call this problem, the ‘small’ measurement problem).

6 SeeBrown (2005) for a different view about the role of dynamical analyses in special relativity;
and Brown and Timpson (2007) for a criticism of the analogy with special relativity.
342 M. Hemmo and O. Shenker

What is the solution they propose to the consistency problem in quantum

mechanics? They appeal to decoherence theory which analyses quantum mechani-
cally the interaction between macroscopic systems and the environment.7 Consider
a typical quantum measurement situation in which an observer (call her Alice)
measures (say, by means of a Stern-Gerlach device; omitted in our description), the
z-spin of an electron, which is prepared in some superposition of the up- and down-
eigenstates of the z-spin. At the ‘initial’ time t0 the quantum state of this universe is
this8 :
√    
|
(t0 ) = 1/ 2 | ↑z + | ↓z ⊗ | ψ 0 (15.1)

where |↑z  and |↓z  are (respectively) the up- and down-eigenstates of the z-spin of
the electron; and | ψ0  is Alice’s brain state in which she is ready to measure the
z-spin of the electron. The Schrödinger equation (in the non-relativistic case) maps
the state in (1) to the following superposition state at the end of the measurement at
some time t:
√   √  
|
(t) = 1/ 2 | ↑z ⊗ |ψ ↑ + 1/ 2 | ↓z ⊗ |ψ ↓ (15.2)

where the | ψ↑↓  are the brain states of Alice in which she would believe that the
outcome of the z-spin measurement is up (or down, respectively), if the initial state
of the electron were only the up-state (or only the down-state). We assume here that
Alice’s brain states | ψ↑↓  include all the degrees of freedom that are correlated with
whatever macroscopic changes occurred in Alice’s brain due to this transition. For
simplicity, the description in (2) omits many details of the microphysics, but nothing
in our argument hinges on this. This is the generic case in which the measurement
problem arises in which Alice’s brain states and the spin states of the electron are
entangled.
The essential idea of decoherence (for our purposes) is very briefly this. Since
Alice is a relatively massive open system that continuously interacts with a large
number of degrees of freedom in the environment, the quantum state of the universe
is not (2), but rather this state:

√   √  
|
t  = 1/ 2 | ↑z ⊗ |ψ ↑ ⊗ |E+ (t) + 1/ 2 |↓z ⊗ |ψ ↓ ⊗ |E− (t) (15.3)

in which the | E± (t) are the collective (relative-)states of all the degrees of freedom
in the environment9 . According to decoherence theory, the Hamiltonian that governs

7 For decoherence theory, and references, see e.g., Joos et al. (2003).
8 Interactionsnever begin, so state (1) (which is a product state) should be replaced with a state in
which Alice + electron are (weakly) entangled; otherwise the dynamics would not be reversible.
9 In discrete models, both | ψ  and | E (t) are states of the form:
N 
↑↓ ± i=1 | μi defined in the
corresponding multi-dimensional Hilbert spaces.
15 Quantum Mechanics as a Theory of Probability 343

the interaction with the environment commutes (or approximately commutes) with
some observables of the macroscopic system, for example the position of its center
of mass. In our example, Alice’s brain states | ψ↑↓  denote the eigenstates of the
observables on which the interaction Hamiltonian depends. These states are called
in the literature the ‘decoherence basis’. The interaction couples (approximately)
the states of the environment | E± (t) to the states | ψ↑↓  in the decoherence basis.
Since the environment consists of a large number of dynamically independent
degrees of freedom (this is expressed by assuming a suitable ‘random’ probability
distribution over the states of the different degrees of freedom in the environment),
the | E± (t) are (highly likely to become) extremely quickly almost orthogonal to
each other. Because of the large number of degrees of freedom in the environment,
this behaviour is likely to persist for very long times (in some models, for times
comparable with the age of the universe) for exactly the same reasons.
As long as decoherence persists, that is the | E± (t) remain nearly orthogonal
to each other, the interference terms in the superposition (3) in the decoherence
basis are very small. This means that at all these times, the off-diagonal elements of
Alice’s reduced density matrix (calculated by partial tracing over the environmental
degrees of freedom) in the decoherence basis | ψ↑↓  are very small. Since the
interference small the branches defined in the decoherence basis | ψ↑↓  evolve in
time in the course of the Schrödinger evolution of (3) almost separately of each
other. That is, given the Schrödinger evolution of the universal state (3), which
results in decoherence, it turns out that (in the decoherence basis) each of the
branches in (3) evolves in a way that is almost independent of the other branches.
In these senses, the decoherence basis is special: It has the above formal features
when decoherence obtains, were in other bases in which state (3) may be written,
the expansion is much more cumbersome.
Notice that in our outline of decoherence we have defined the branches by Alice’s
brain states; but this is just an artefact of our example. The states of any decohering
system (such as a macroscopic body in a thermal or radiation bath, a Brownian
particle in suspension in a fluid, a pointer of a measuring device, Schrödinger’s cat)
are equally suitable to define the branches. It might even be that Alice’s relevant
brain states are not subject directly to decoherence (whether or not this is the case
is the subject matter of brain science after all). All one needs is that Alice’s relevant
brain states (whatever they are) that give rise to her experience be coupled to the
states of a decohering system. In state (3) we use Alice’s brain states to define the
decoherence basis only for simplicity.
The essential point for Bub and Pitowsky is that, since the interference terms are
very small in the decoherence basis, the dynamical process of decoherence gives rise
to an effectively classical (Boolean) probability space in which the probabilities for
the events are given by the Born probabilities. The macroscopic events in this space
are given by the projections onto the eigenstates of the decohering variables (in our
example, the projections onto the states in the decoherence basis | ψ↑↓ ) and these
turn out to be the macroscopic variables that feature in our experience, and more
generally in classical mechanics. And the probability space (approximately) satisfies
classical probability theory as long as decoherence endures. For example, the
344 M. Hemmo and O. Shenker

conditional probabilities defined by sequences of such Boolean spaces (associated

with sequences of measurements) are additive. Since the process of decoherence
is a consequence of the unitary dynamics, Bub and Pitowsky take this analysis to
provide the consistency proof they seek, that is, a proof that shows how a classical
probability space of the kind of events we experience emerges in decoherence
situations, without (allegedly) adding anything to Hilbert space.
‘The analysis shows that a quantum dynamics, consistent with the kinematics
of Hilbert space, suffices to underwrite the emergence of a classical probability
space for the familiar macroevents of our experience, with the Born probabilities
for macroevents associated with measurement outcomes derived from the quantum
state as a credence function. The explanation for such non-classical effects as the
loss of information on conditionalization is not provided by the dynamics, but by
the kinematics, and given ‘no cloning’ as a fundamental principle, there can be no
deeper explanation. In particular, there is no dynamical explanation for the definite
occurrence of a particular measurement outcome, as opposed to other possible
measurement outcomes in a quantum measurement process – the occurrence is
constrained by the kinematic probabilistic correlations encoded in the projective
geometry of Hilbert space, and only by these correlations.’ (Bub and Pitowsky, p.
454).
However, this doesn’t work. In special relativity ‘consistency proofs’ show that
Lorentz covariant forces ‘track’ the structure of Minkowski spacetime of which
Lorentz contraction and time dilation are real physical features. So, by the analogy,
the quantum consistency proof should prove – without adding structure to Hilbert
space – that decoherence gives rise to the emergence of an effectively classical
probability space of the events in our experience and that the von-Neumann-Lüders
conditionalisation ‘tracks’ real physical features of events occurring in the universe
(the objective chances). But we shall argue that this cannot be done. The analogy
with special relativity breaks down, because even in decoherence situations the
quantum (unitary) dynamics of the universe does not result in the emergence of
an effectively classical (Boolean) probability space. In the next section we explain
why.

15.3 The Role of Decoherence in the Consistency Proof

It is true that the decoherence process is basis-independent and that the interaction
Hamiltonian picks out the decoherence basis as dynamically preferred (given a
factorisation of the Hilbert space into system and environment), as we explained
above. In fact, since the process of decoherence is continuous, there is a continuous
infinity of basis states in Alice’s state space other than the | ψ↑↓  which satisfy to
the same degree the conditions of decoherence (namely the near orthogonality of
the relative states of the environment and the fact that this behaviour is stable over
long time periods). It is an objective fact that as long as decoherence endures the
interference between the branches defined in any such basis (and the off-diagonal
15 Quantum Mechanics as a Theory of Probability 345

elements in the reduced state of the decohering system) are very small and can
hardly be detected by measurements. In many cases in physics it is practically useful
to ignore small terms which make no detectible difference for calculating empirical
predictions. Here are two examples.
In probability theory it is sometimes assumed that events with very small
probability will not occur with practical (or moral) certainty, or equally, that events
with very high probability which is not strictly one, will occur with practical
certainty (see Kolmogorov 1933). In these cases, the identification of very small
probability with zero probability is practically useful whenever one assumes that
cases with very small probability occur with very small relative frequency. Although
such cases will appear in the long run of repeated trials, they are unlikely to appear
in our relatively short-lived experience. Another example is the Taylor series of
a function which is used to approximate the value of a function at the expansion
point by the function’s derivatives at that point (when the derivatives are defined
and continuous around the point). Although the Taylor series is an infinite sum of
terms, it is true in many cases (but not always) that one can approximate the function
to a sufficient degree of accuracy by truncating the Tailor series and taking only
the first finite number of the terms in the series. So effectively in these cases the
rest of the terms in the series which typically decay as the order of the derivatives
increases, but do not strictly vanish, are set to zero. One can obtain a bound via
Taylor’s theorem, on the size of the error in the approximation depending on the
length of the truncated series which might be much below the degree of accuracy
that one can measure. Although in general the Tailor infinite series is not equal
to the function (even if the series converges at every point), using the truncated
series gives accurate empirical predictions about the values of the function near the
expansion points and its behavior in various limits. But no claim is made that the
truncated small terms do not exist, the claim is only that we are not able to detect
their existence in our experience.
It seems to us that the role of decoherence in the foundations of quantum
mechanics should be understood along similar lines to the above two examples.
Suppose that the Hamiltonian of the universe includes interactions that we call
measurements of certain observables and decoherence interactions (as in our spin
example above). The evolution of the universal state can be written in infinitely
many bases in the Hilbert space of the universe. Different bases pick out different
events, in some bases the interference terms between the components describing
the events are small, and in others they are large. In some intuitive sense, the bases
in which the interference terms are small seem to be esthetically more suitable or
more convenient for us to describe the quantum state. But given the Hilbert space
structure these bases have no other special status, in particular they are not more
physically real (as it were) than the other bases. In this sense the analogy of Hilbert
space bases with Lorentz frames in special relativity is in place. It turns out that
the events that feature in our experience correspond to the projections onto the
decoherence basis. If we take this fact into account without a deeper explanation
of how it comes about, then it is only natural that we write the quantum state
describing our measurements in the decoherence basis. And then for matters of
346 M. Hemmo and O. Shenker

calculating our empirical predictions the small interference terms in the decoherence
basis make no detectable difference to what we can measure in our experience, and
for this purpose these terms can be set to zero. To detect the superposition in (3), by
means of measurements which are described in the decoherence basis, one needs to
manipulate the relative states of the environment in a way that results in significant
interference. But this is practically impossible, due to the large number of degrees
of freedom in the environment and the fact that they are much beyond our physical
reach. In this practical sense, setting the interference terms in the decoherence basis
to effectively zero is a perfectly useful approximation.
But in order for Bub and Pitowsky’s consistency proof to be analogous to the
dynamical proofs in special relativity, these effective considerations are immaterial.
It is an objective fact about the universe that decoherence obtains (given a factori-
sation). This is our candidate (iii) for the analogy with the breaking of the thread in
special relativity. But what about Bub and Pitowsky’s target (candidate (ii)) for the
analogy, namely the fact that the structure of the probability space of the events in
our experience seems to be classical (Boolean)? The decoherence basis is analogous
only to certain Lorentz frames, and the decoherence interaction explains why the
interference terms of state (3), described in this basis, are small; the decoherence
interaction equally explains why in other bases of the Hilbert space of the universe
(which are analogous to other Lorentz frames) the interference terms of state (3) are
large. Consequently, the decoherence interaction explains why a probability space
of events picked out by the decoherence basis is effectively classical (Boolean);
the decoherence interaction equally explains why the probability space of events
picked out by some other bases are not effectively classical. But the decoherence
interaction does not explain why events picked out by the decoherence basis, rather
than by these other bases, feature in our experience.
For this reason, the analogy with special relativity breaks down. In special
relativity one can explain why the experiences of certain (physical) observers are
associated with certain Lorentz frames by straightforward physical facts about these
observers, for example, their relative motion.10 But in quantum mechanics, if we
do not add structure to Hilbert space, one cannot explain by appealing to physical
facts why the experiences of certain observers, for example, human observers, are
associated with the events picked out by the decoherence basis. The emergence of
the probability space of these events in our experience is not a basis-independent

10 In statistical mechanics, here is an analogous scenario: the phase space of the universe can
be partitioned into infinitely many sets corresponding to infinitely many macrovariables. Some
of these partitions exhibit certain regularities, for example, the partition to the thermodynamic
sets which human observers are sensitive to, exhibit the thermodynamic regularities; while other
partitions to which human observers are not sensitive, although they are equally real exhibit
other regularities and may even be anomalous (that is, exhibit no regularities at all). We call this
scenario in statistical mechanics Ludwig’s problem (see Hemmo and Shenker 2012, 2016). But
one can explain by straightforward physical facts why (presumably) human observers experience
the thermodynamic sets and regularities but not the equally existing other sets (although it might
be that in our experience some other sets also appear). This case too is dis-analogous to the basis-
symmetry in quantum mechanics, from which the problem of no preferred basis follows.
15 Quantum Mechanics as a Theory of Probability 347

fact in quantum mechanics. And if quantum mechanics is a complete theory, as Bub

and Pitowsky assume, then it follows that components of the universal state in some
or other basis do not pick out anything that is physically ‘more real’ (as it were) than
components in any other basis. In particular, the decoherence basis has absolutely
no preferred ontological status by comparison to any other basis in the Hilbert space
of the universe, despite the fact that it is special in that the interference terms of
states like (3) described in this basis are small. So: what is the explanation of the
fact that events picked out by the decoherence basis (for which the probability space
is effectively classical) feature in our experience? On the Bub-Pitowsky approach
this cannot be explained by appealing to facts about, say, Alice’s brain and its
interactions, for the same reason that (as we see from our example), Alice’s brain,
even if it is subject to decoherence, can be equally described by infinitely many
bases all of which are on equal status. The only way to explain why we experience
events that are picked out by the decoherence basis, rather than events picked out by
other bases, is to add laws, or structure, to Hilbert space, from which it will follow
that the decoherence basis is physically preferred, in the sense that it picks out the
actually occurring events in our experience.11
What about the quantum conditionalisation (and von Neumann’s projection
postulate)? One might argue that the von Neumann-Lüders conditionalisation does
not amount to adding structure to Hilbert space, since it expresses the information
loss in measurement, which is itself entailed by the ‘no cloning’ theorem. But this
is beside the point, since the conditionalisation depends not only on the individual
outcome that actually occurs (a desideratum that Bub and Pitowsky do not attempt
to explain), but also on a choice of a Hilbert space basis. As we argued above, the
Hilbert space structure, however, says nothing about the preference, or choice, of
any basis, let alone the occurrence (or non-occurrence) of individual events. This
is why the additional laws (or structure) are required. This is why decoherence
cannot provide the consistency proof Bub and Pitowsky seek by their analogy with
relativity theory.

15.4 From Subjectivism to Idealism

In the Bayesian approach to probability, probability theory is about how to make

rational decisions in the face of uncertainty, and therefore the starting point is
that probability is a measure of ignorance (a credence function describing degrees
of belief) in a single trial, and no a priori connection is assumed between the
probability and limiting frequencies, except as derived by conditionalisation on the
experienced frequencies, and adding inductive generalisations.

11 A preferred basis in exactly the same sense is also required in all the contemporary versions of
the Everett (1957) interpretation of quantum mechanics, so that one must add a new law of physics
to the effect that worlds emerge relative to the decoherence basis; see (Hemmo and Shenker 2019).
348 M. Hemmo and O. Shenker

In the face of the measurement problem, the Bayesian approach to quantum

mechanics says this.12 The quantum state is just a credence function, tracking the
limiting frequencies in experience, it is not a bookkeeping device for tracking the
‘universe’s objective chances’ (as it is in the Bub-Pitowsky approach). The credence
function changes in accordance with the unitary dynamics in the absence of mea-
surements, but in measurement contexts it changes by quantum conditionalisation
(the projection postulate or the von Neumann-Lüders rule), which is the quantum
analogue of the classical Bayesian updating. Both laws are seen as inductive
generalisations from experience. Decoherence in this approach is not a physical
process but some sort of a mental process that is subject to a rationality constraint
on the agent’s experiences, in particular, belief states, which is related to so-called
Dutch-book consistency.13 In this approach no attempt is made at explaining how
our experience comes about.
On this approach quantum mechanics is a theory about the mental states
(experiences, beliefs) of observers. Each observer describes the rest of the world
by the unitary dynamics, and applies the quantum conditionalisation law only
to the evolution of its own mental state. This extreme form of subjectivism is
necessary to avoid contradictions in the ‘internal scenarios’ (or the experience) of
each observer.14 This is because, if there are two observers or more, nothing prevents
a case in which different observers will experience different outcomes for the same
measurement. But this does not lead to a contradiction in the internal scenario (as
it were) of each observer separately: as is well known, quantum mechanics ensures
that it will always seem to me that the experiences reported by other observers are
the same as mine. On this picture, quantum mechanics is not about anything external
to experience. Whether or not the experience is somehow correlated with anything
external (and whether there is anything external) is beyond quantum mechanics.
We wish to point out two consequences of this approach. First, as in the Bub-
Pitowsky approach, the Bayesian approach adds structure to Hilbert space. In
implementing the quantum conditionalisation law, experience plays two crucial
roles. The first role of experience is to choose one basis in the Hilbert space of
the measured system, the basis formed by the set of eigenstates of the observable
measured, with respect to which the agent applies the von Neumann-Lüdres rule15 ;
and the second role is to choose one of the eigenstates in this basis as the experienced

12 Recent variations are, for example, Caves et al. (2002a, b, 2007); Fuchs and Schack (2013);
Fuchs et al. (2014); Fuchs and Stacey (2019).
13 For the role of decoherence in the Bayesian approach and its relation to Dutch-book consistency,

see Fuchs and Scack (2012).

14 See Hagar and Hemmo (2006); Hagar (2003).
15 Bohm’s (1952) theory and the GRW theory (Ghirardi et al. 1986) assume a preferred basis

from the start. Our point is that the structure added by the Bayesian approach is no less than
e.g., Bohmiam trajectories or GRW flashes.
15 Quantum Mechanics as a Theory of Probability 349

measurement outcome, where the probability for the choice is dictated by the Born
rule.16
The second consequence of the Bayesian approach is that as a matter of principle,
not of ignorance about the external physical world, mental states (and experience in
general) cannot be accounted for by physics. Here is why. Contemporary science
(brain and cognitive science) is far from explaining how mental experience comes
about. But the standard working hypothesis in these sciences – which is fruitful
and enjoys vast empirical support – is that our mental experience is to be explained
(somehow, by and large) by features of our brains, perhaps together with some other
parts of the body; for simplicity we call all of them ‘the brain’. How and why we
can treat biological systems in effectively classical terms is part of the questioin of
the classical limit of quantum mechanics. But here our specific question is this. Let
us suppose that quantum mechanics does yield in some appropriate limit classical
behavior of (say) our brains. Can the Bayesian approach accept the standard working
hypothesis that mental states are to be accounted for by brain states (call it the ‘brain
hypothesis’)?
If the brain hypothesis is true, there is some physical state of my brain which
underlies my credence function. Since according to contemporary physics quantum
mechanics is our best physical theory, this physical state (whatever its details are)
should be some quantum state of my brain. But on the Bayesian approach, the
quantum state of my brain is a credence function of some observer, not the physical
state of anything; so a fortiori it can’t be the physical state of my brain. It follows
that, as a matter of principle, mental states cannot be accounted for by brain states:
they are not physical even fundamentally.
In classical physics the Bayesian theory of probability is not meant to replace
the physical theory (i.e., classical relativistic physics) describing how the relative
frequencies of events are brought about. After all, as Pitowsky (2003, p. 400) has
put it: “the theory of probability, even in its most subjective form, associates a
person’s degree of belief with the objective possibilities in the physical world.”
In the classical case the subjective interpretation of probability (de Finetti 1970;
Ramsey 1929) may assume in the background that – as a matter of principle –
not only the frequencies but also the credences of agents and their evolution over
time might be ultimately accounted for by physics. In this sense, the subjectivist
approach to classical probability is consistent with the physicalist thesis in cognition
and philosophy of mind. But the Bayesian approach to quantum mechanics, is not
only inconsistent with physicalism: it is a form of idealism.

Acknowledgement We thank Guy Hetzroni and Cristoph Lehner for comments on an earlier draft
of this paper. This research was supported by the Israel Science Foundation (ISF), grant number
1114/18.

16 These two roles are carried out by the mind at one shot (as it were), but analytically they are
different.
350 M. Hemmo and O. Shenker

References

Bell, J. S. (1987a). Are there quantum jumps? In J. Bell (Ed.), Speakable and unspeakable in
quantum mechanics (pp. 201–212). Cambridge: Cambridge University Press.
Bell, J. S. (1987b). How to teach special relativity. In J. Bell (Ed.), Speakable and unspeakable in
quantum mechanics (pp. 67–80). Cambridge: Cambridge University Press.
Brown, H. (2005). Physical relativity: Spacetime structure from a dynamical perspective. Oxford:
Oxford University Press.
Brown, H. R., & Timpson, C. G. (2007). Why special relativity should not be a template for a
fundamental reformulation of quantum mechanics. In Pitowsky (Ed.), Physical theory and its
interpretation: Essays in honor of Jeffrey Bub, W. Demopoulos and I. Berlin: Springer.
Bub, J. (1977). Von Neumann’s projection postulate as a probability conditionalization rule in
quantum mechanics. Journal of Philosophical Logic, 6, 381–390.
Bub, J. (2007). Quantum probabilities as degrees of belief. Studies in History and Philosophy of
Modern Physics, 38, 232–254.
Bub, J. (2016). Bananaworld: Quantum mechanics for primates. Oxford: Oxford University Press.
Bub, J. (2020). ‘Two Dogmas’ Redux. In Hemmo, M., Shenker, O. (eds.) Quantum, probability,
logic: Itamar Pitowsky’s work and influence. Cham: Springer.
Bub, J., & Pitowsky, I. (2010). Two dogmas about quantum mechanics. In S. Saunders, J. Barrett,
A. Kent, & D. Wallace (Eds.), Many worlds? Everett, quantum theory, and reality (pp. 431–
456). Oxford: Oxford University Press.
Bohm, D. (1952). A suggested interpretation of the quantum theory in terms of “hidden” variables.
Physical Review, 85(166–179), 180–193.
Caves, C. M., Fuchs, C. A., & Schack, R. (2002a). Unknown quantum states: The quantum de
Finetti representation. Journal of Mathematical Physics, 43(9), 4537–4559.
Caves, C. M., Fuchs, C. A., & Schack, R. (2002b). Quantum probabilities as Bayesian probabilities.
Physical Review A, 65, 022305.
Caves, C. M., Fuchs, C. A., & Schack, R. (2007). Subjective probability and quantum certainty.
Studies in History and Philosophy of Modern Physics, 38, 255–274.
de Finetti, B. (1970). Theory of probability. New York: Wiley.
Everett, H. (1957). ‘Relative state’ formulation of quantum mechanics. Reviews of Modern Physics,
29, 454–462.
Fuchs C. A. & Schack, R. (2012). Bayesian conditioning, the reflection principle, and quantum
decoherence. In Ben-Menahem, Y., & Hemmo, M. (eds.), Probability in physics (pp. 233–247).
The Frontiers Collection. Berlin/Heidelberg: Springer.
Fuchs, C. A., & Schack, R. (2013). Quantum Bayesian coherence. Reviews of Modern Physics, 85,
1693–1715.
Fuchs, C. A., Mermin, N. D., & Schack, R. (2014). An introduction to QBism with an application
to the locality of quantum mechanics. American Journal of Physics, 82, 749–754.
Fuchs, C. A., & Stacey, B. (2019). QBism: Quantum theory as a hero’s handbook. In E. M.
Rasel, W. P. Schleich, & S. Wölk (Eds.), Foundations of quantum theory: Proceedings of the
International School of Physics “Enrico Fermi” course 197 (pp. 133–202). Amsterdam: IOS
Press.
Ghirardi, G., Rimini, A., & Weber, T. (1986). Unified dynamics for microscopic and macroscopic
systems. Physical Review, D, 34, 470–479.
Hagar, A. (2003). A philosopher looks at quantum information theory. Philosophy of Science,
70(4), 752–775.
Hagar, A., & Hemmo, M. (2006). Explaining the unobserved – Why quantum mechanics ain’t only
about information. Foundations of Physics, 36(9), 1295–1324.
Hemmo, M., & Shenker, O. (2012). The road to Maxwell’s demon. Cambridge: Cambridge
University Press.
Hemmo, M., & Shenker, O. (2016). Maxwell’s demon. Oxford University Press: Oxford Hand-
books Online.
15 Quantum Mechanics as a Theory of Probability 351

Hemmo, M. & Shenker, O. (2019). Why quantum mechanics is not enough to set the framework
for its interpretation, forthcoming.
Joos, E., Zeh, H. D., Giulini, D., Kiefer, C., Kupsch, J., & Stamatescu, I. O. (2003). Decoherence
and the appearance of a classical world in quantum theory. Heidelberg: Springer.
Pitowsky, I. (2003). Betting on the outcomes of measurements: A Bayesian theory of quantum
probability. Studies in History and Philosophy of Modern Physics, 34, 395–414.
Kolmogorov, A. N. (1933). Foundations of the Theory of Probability. New York: Chelsea
Publishing Company, English translation 1956.
Pitowsky, I. (2007). Quantum mechanics as a theory of probability. In W. Demopoulos & I.
Pitowsky (Eds.), Physical theory and its interpretation: Essays in honor of Jeffrey Bub (pp.
213–240). Berlin: Springer.
Ramsey, F. P. (1990). Truth and Probability. (1926); reprinted in Mellor, D.H. (ed.), F. P. Ramsey:
Philosophical Papers. Cambridge: Cambridge University Press.
Tumulka, R. (2006). A relativistic version of the Ghirardi–Rimini–Weber model. Journal of
Statistical Physics, 125, 821–840.
von Neumann, J. (2001). Unsolved problems in mathematics. In M. Redei & M. Stoltzner (Eds.),
John von Neumann and the foundations of quantum physics (pp. 231–245). Dordrecht: Kluwer
Academic Publishers.
Chapter 16
On the Three Types of Bell’s Inequalities

Gábor Hofer-Szabó

Abstract Bell’s inequalities can be understood in three different ways depending

on whether the numbers featuring in the inequalities are interpreted as classical
probabilities, classical conditional probabilities, or quantum probabilities. In the
paper I will argue that the violation of Bell’s inequalities has different meanings in
the three cases. In the first case it rules out the interpretation of certain numbers as
probabilities of events. In the second case it rules out a common causal explanation
of conditional correlations of certain events (measurement outcomes) conditioned
on other events (measurement settings). Finally, in the third case the violation
of Bell’s inequalities neither rules out the interpretation of these numbers as
probabilities of events nor a common causal explanation of the correlations between
these events – provided both the events and the common causes are interpreted non-
classically.

Keywords Bell’s inequalities · Conditional probability · Correlation polytope

16.1 Introduction

Ever since its appearance three decades ago, Itamar Pitowsky’s Quantum Probabil-
ity – Quantum Logic has been serving as a towering lighthouse showing the way
for many working in the foundations of quantum mechanics. In this wonderful book
Pitowsky provided an elegant geometrical representation of classical and quantum
probabilities: a powerful tool in tackling many difficult formal and conceptual
questions in the foundations of quantum theory. One of them is the meaning of
the violation of Bell’s inequalities.
However, for someone reading the standard foundations of physics literature on
Bell’s theorems, it is not easy to connect up Bell’s inequalities as they are presented

G. Hofer-Szabó ()
Research Center for the Humanities, Budapest, Hungary
e-mail: szabo.gabor@btk.mta.hu

© Springer Nature Switzerland AG 2020 353

M. Hemmo, O. Shenker (eds.), Quantum, Probability, Logic,
Jerusalem Studies in Philosophy and History of Science,
https://doi.org/10.1007/978-3-030-34316-3_16
354 G. Hofer-Szabó

in Pitowsky’s book with the inequalities presented by Bell, Clauser, Horne, Shimony
etc. The main difference, to make it brief, is that Bell’s inequalities in their original
form are formulated in terms of conditional probabilities, representing certain
measurement outcomes provided that certain measurement are performed, while
Pitowsky’s Bell inequalities are formulated in terms of unconditional probabilities,
representing the distribution of certain underlying properties or events responsible
for the measurement outcomes.
Let us see this difference in more detail. Probabilities enter into quantum
mechanics via the trace formula

p = Tr(ρ̂ Â) (16.1)

where ρ̂ is a density operator, Â is the spectral projection associated to eigenvalue α

of the self-adjoint operator â, and Tr is the trace function. Let us call the probabilities
generated by the trace formula (16.1) quantum probabilities.
Now, what is the physical interpretation of quantum probabilities? There are
two possible answers to this question according to two different interpretations of
quantum mechanics:
1. On the operational (minimal) interpretation, the density operator ρ̂ represents
the state or preparation s of the system; the self-adjoint operator â represents
the measurement a performed on the system; and the spectral projection Â
represents the outcome A of the measurement a. On this interpretation the
quantum probability is interpreted as the conditional probability

p = ps (A|a) (16.2)

that is the probability of obtaining outcome A provided that the measurement

a has been performed on the system previously prepared in state s. This
interpretation is called minimal since the fulfillment of (16.2) is a necessary
condition for the theory to be empirically adequate.
2. On the ontological (deterministic hidden variable, property) interpretation, the
density operator ρ̂ represents the distribution ρ of the ontological states λ ∈ 
in the preparation s; the operator â represents the physical magnitude a ∗ ; and
the projection  represents the event A∗ that the value of a ∗ is A∗ .1 Denote
∗ ∗
by A = {λA } the set of those ontological states for which the value a ∗ is
∗
A . On the ontological interpretation the quantum probability is (intended to be)
interpreted as the unconditional probability

p = ps (A∗ ) = ρ(λ) dλ (16.3)
∗
A

that is the probability of the event A∗ in the state s.

1I denote the event and the value by the same symbol.

16 On the Three Types of Bell’s Inequalities 355

The ontological and the operational definition are connected as follows. The
measurement a is said to measure the physical magnitude a ∗ only if the following
holds: the outcome of measurement a performed on the system will be A if and
only if the system is in an ontological state for which the value of a ∗ is A∗ :
∗
pλ (A|a) = 1 if and only if λ ∈ A

As is seen, the two interpretations differ in how they treat quantum probabilities.
On the operational interpretation quantum probabilities are condition probabilities,
while on the ontological interpretation they are unconditional probabilities. (Note
that in both interpretations probabilities are “conditioned” on the preparation which
will be dropped from the next section.)
Now, consider a set of self-adjoint operators {âi } (i ∈ I ) each with two spectral
projections {Âi , Â⊥
i }. Correspondingly, consider a set of measurements {ai } each
with two outcomes {Ai , A⊥ ∗
i } and a set of magnitudes {ai } each with two values
{A∗i , A∗⊥
i }. If Âi and Âj are commuting, then the quantum probabilities

pi = Tr(ρ̂ Âi ) (16.4)

pij = Tr(ρ̂ Âi Âj ) (16.5)

can be interpreted either operationally:

pi = ps (Ai |ai ) (16.6)

pij = ps (Ai ∧ Aj |ai ∧ aj ) (16.7)

or ontologically:

pi = ps (A∗i ) (16.8)
pij = ps (A∗i ∧ A∗j ) (16.9)

Consider a paradigmatic Bell inequality, the Clauser-Horne inequalities:

−1  pij + pi  j + pij  − pi  j  − pi − pj  0
i, i  = 1, 2; j, j  = 3, 4; i = i  ; j = j  (16.10)

The numbers featuring in (16.10) are probabilities. But which type of probabilities?
Are they simply (uninterpreted) quantum probabilities of type (16.4)–(16.5), or
condition probabilities of type (16.6)–(16.7), or unconditional probabilities of
type (16.8)–(16.9)? Depending on how the probabilities in the Clauser-Horne
inequalities are understood, the inequalities can be written out in the following three
different forms:
356 G. Hofer-Szabó

− 1  Tr(ρ̂ Âi Âj ) + Tr(ρ̂ Âi  Âj ) + Tr(ρ̂ Âi Âj  )

−Tr(ρ̂ Âi  Âj  ) − Tr(ρ̂ Âi ) − Tr(ρ̂ Âj )  0 (16.11)

(where Âi Âj , Âi  Âj , Âi Âj  and Âi  Âj  are commuting projections); or

− 1  ps (Ai ∧ Aj |ai ∧ aj ) + ps (Ai  ∧ Aj |ai  ∧ aj ) + ps (Ai ∧ Aj  |ai ∧ aj  )

−ps (Ai  ∧ Aj  |ai  ∧ aj  ) − ps (Ai |ai ) − ps (Aj |aj )  0 (16.12)

or

− 1  ps (A∗i ∧ A∗j ) + ps (A∗i  ∧ A∗j ) + ps (A∗i ∧ A∗j  )

−ps (A∗i  ∧ A∗j  ) − ps (A∗i ) − ps (A∗j )  0 (16.13)

Thus, altogether we have three different types of Bell inequalities depending on

whether and how the probabilities featuring in (16.10) are physically interpreted.
The aim of the paper is to clarify as to what is exactly excluded if Bell’s inequalities
of type (16.11), (16.12) or (16.13) are violated. I will argue that the violation has
three different meanings in the three different cases. In case of (16.13), when the
probabilities are classical unconditional probabilities, it rules out the interpretation
of certain numbers as probabilities of events or properties. These are the inequalities
which Pitowsky identified and categorized. The violation of inequalities (16.12),
when the probabilities are classical conditional probabilities, does not rule out the
interpretation of certain numbers as conditional probabilities of events but only a
common causal explanation of the conditional correlations between these events.
These are the Bell inequalities used in the standard foundations of physics literature.
Finally, the violation of Bell’s inequalities (16.11) neither rules out the interpretation
of certain numbers as probabilities of events nor a common causal explanation of the
correlations between these events—provided that both events and common causes
are interpreted non-classically. The violation of these Bell’s inequalities is used for
another purpose: it places a bound on the strength of correlations between these
events.
I will proceed in the paper as follows. In Sect. 16.2 I analyze Bell’s inequalities
for classical probabilities, and in Sect. 16.3 for classical conditional probabilities.
In Sect. 16.4 the two types will be compared. I turn to Bell’s inequalities in terms
of quantum probabilities in Sect. 16.5. In Sect. 16.6 I apply the results to the EPR-
Bohm scenario and finally conclude in Sect. 16.7.
To make the notation simple, I drop both the hat and the asterisk from the next
section on, that is I write A instead of both  and also A∗ . The semantics of A
will be clear from the context: in classical conditional probabilities A will refer to
a measurement outcome, in classical unconditional probabilities it will refer to a
property/event, and in quantum probabilities it will refer to a projection.
16 On the Three Types of Bell’s Inequalities 357

My approach strongly relies on Szabó (2008) distinction between Bell’s original

inequalities and what he calls the Bell-Pitowsky inequalities. For a somewhat
parallel research see Gömöri and Placek (2017).

16.2 Case 1: Bell’s Inequalities for Classical Probabilities

Consider real numbers pi and pij in [0, 1] such that i = 1 . . . n and (i, j ) ∈ S
where S is a subset of the index pairs {(i, j ) |i < j ; i, j = 1 . . . n }. When can these
numbers be probabilities of certain events and their conjunctions? More precisely:
given the numbers pi and pij , is there a classical probability space (, , p) with
events Ai and Ai ∧ Aj in  such that

pi = p(Ai )
pij = p(Ai ∧ Aj )

In brief, do the numbers pi and pij admit a Kolmogorovian representation?

Itamar Pitowsky’s (1989) provided an elegant geometrical answer to this ques-
tion. Arrange the numbers pi and pij into a vector
 
p = p1 , . . . , pn ; . . . , pij , . . .

called correlation vector. p is an element of an n+|S| dimensional real linear space,

R(n, S) ∼= Rn+|S| , where |S| is the cardinality of S.
Now, we construct a polytope in R(n, S). Let ε ∈ {0, 1}n . To each ε assign a
classical vertex vector (truth-value function): u ε ∈ R(n, S) such that

uiε = εi i = 1...n
uijε = εi ε j (i, j ) ∈ S

Then we define the classical correlation polytope in R(n, S) as the convex hull of
the classical vertex vectors:
⎧  ⎫
⎨    ⎬

c(n, S) := p ∈ R(n, S) p = λε u ε ; λε  0; λε = 1
⎩  ⎭
ε∈{0,1}n ε∈{0,1}n

The polytope c(n, S) is a simplex, that is any correlation vector in c(n, S) has a
unique expansion by classical vertex vectors.
Pitowsky’s theorem (Pitowsky 1989, p. 22) states that p admits a Kolmogorovian
representation if and only if p ∈ c(n, S). That is numbers can be probabilities of
certain events and their conjunctions if and only if the correlation vector composed
of these numbers is in the classical correlation polytope.
358 G. Hofer-Szabó

Now, Bell’s inequalities enter the scene as the facet inequalities of the classical
correlation polytopes. The simplest such correlation polytope is the one with n =
2 and S = {(1, 2)}. In this case the vertices are: (0, 0; 0), (1, 0; 0), (0, 1; 0) and
(1, 1; 1) and the facet inequalities (Bell’s inequalities) are the following:

0  p12  p1 , p2  1
p1 + p2 − p12  1

Another famous polytope is c(n, S) with n = 4 and S = {(1, 3), (1, 4), (2, 3),
(2, 4)}. The facet inequalities are then the following:

0  pij  pi , pj  1 i = 1, 2; j = 3, 4 (16.14)
pi + pj − pij  1 i = 1, 2; j = 3, 4 (16.15)
−1  pij +pi  j +pij  −pi  j  −pi −pj  0 i, i =1, 2; j, j =3, 4; i = i ; j = j 
  

(16.16)

The facet inequalities (16.16) are called the Clauser-Horne inequalities. They
express whether 4 + 4 real numbers can be regarded as the probability of four
classical events and their certain conjunctions.
A special type of the correlation vectors are the independence vectors that is
correlation vectors such that for all (i, j ) ∈ S, pij = pi pj . It is easy to see that all
independence vectors lie in the classical correlation polytope with coefficients

%
n 
pi if ε i = 1
λε = pi∗ , where pi∗ = (16.17)
1 − pi if ε i = 0
i=1

Classical vertex vectors are independence vectors by definition; they are extremal
points of the classical correlation polytope. For a classical vertex vector pi ∈
{0, 1} for all i = 1 . . . n. Thus we will sometimes call a classical vertex vector
a deterministic independence vector and an independence vector which is not a
classical vertex vector an indeterministic independence vector.
Although correlation vectors in c(n, S) has a unique convex expansion by
classical vertex vectors, they (if not classical vertex vectors) can have many convex
expansions by indeterministic independence vectors. Moreover, for a classical
correlation vector p ∈ c(n, S) (which is not a classical vertex vector) there always
exist many sets of indeterministic independence vectors {p ε } such that
 
p = λε u ε = λε p ε
ε ε
16 On the Three Types of Bell’s Inequalities 359

that is the coefficients λε of the expansion of p by classical vertex vectors and by

indeterministic independence vectors are the same.2
Let us call a correlation vector which is not an independence vector a proper
correlation vector. For a proper correlation vector p there is at least one pair
(i, j ) ∈ S such that pij = pi pj . When lying in the classical correlation polytope p
represents the probabilities of certain events and their conjunctions and the events
associated to indices i and j are correlated:

p(Ai ∧ Aj ) = p(Ai ) p(Aj ) (16.19)

Now, we ask the following question: When do the correlations in the Kolmogorovian
representation of a proper correlation vector in c(n, S) have a common causal
explanation?
A common cause in the Reichenbachian sense (Reichenbach 1956) is a screener-
off partition of the algebra  (or an extension of the algebra; see Hofer-Szabó et al.
2013, Ch. 3 and 6). In other words, correlations (16.19) are said to have a joint
common causal explanation if there is a partition {Ck } (k ∈ K) of  such that for
any (i, j ) in S and k ∈ K

p(Ai ∧ Aj |Ck ) = p(Ai |Ck ) p(Aj |Ck ) (16.20)

Introduce the notation

pik = p(Ai |Ck )

ck = p(Ck )

where k ck = 1 and construct for each k ∈ K a common cause vector
 
p k = p1k , . . . , pnk ; . . . , pik pjk , . . .

2 For example:
 
2 2 1
p = , ;
5 5 5
1 1 1 2
= (1, 1; 1) + (1, 0; 0) + (0, 1; 0) + (0, 0; 0)
5 5 5 5
  & √ √ √ '
1 1 1 1 1 3+ 5 3+ 5 7+3 5
= , ; + , ;
5 4 4 16 5 8 8 32
& √ √ √ '  
1 3 − 5 3 − 5 7 − −3 5 2 1 1 1
+ , ; + , ; (16.18)
5 8 8 32 5 2 2 4
360 G. Hofer-Szabó

 Due to (16.20) and the theorem of total

for the classical correlation vector p.
probability

p = ck p k
k

Since the common cause vectors are independence vectors lying in the classical
correlation polytope c(n, S), therefore their convex combination p also lies in the
classical correlation polytope—which, of course, we knew since we assumed that p
has a Kolmogorovian representation.
We call a common cause vector p k deterministic, if p k is a deterministic
independence vector (classical vertex vector); otherwise we call it indeterministic.
All classical correlation vectors have 2n deterministic common cause vectors,
namely the classical vertex vectors. In this cases k = ε ∈ {0, 1}n and the
probabilities are:

piε = ε i
cε = λε

where λε is specified in (16.17). Classical correlation vectors which are not classical
vertex vectors also can be expanded as a convex sum of indeterministic common
cause vectors in many different ways.
Conversely, knowing the k common cause vectors and the probabilities ck alone,
one can easily construct a common causal explanation of the correlations (16.19).
First, construct the classical probability space  associated to the correlation vector
p (for the details see Pitowsky 1989, p. 23). Then extend  containing the events
Ai and Ai ∧ Aj such that the extended probability space contains also the common
causes {Ck } (for such an extension see Hofer-Szabó et al. (1999)).
To sum up, in the case of classical probabilities the fulfillment of Bell’s
inequalities is a necessary and sufficient condition for a set of numbers to be the
probability of certain events and their conjunctions. If these events are correlated,
then the correlations will always have a common causal explanation. In other
words, having a common causal explanation does not put a further constraint on
the correlation vectors. Thus, Bell’s inequalities have a double meaning: they test
whether numbers can represent probabilities of events and at the same time whether
the correlations between these events (provided they exist) have a joint common
cause. These two meanings of Bell’s inequalities will split up in the next section
where we treat Bell’s inequalities with classical conditional probabilities.
16 On the Three Types of Bell’s Inequalities 361

16.3 Case 2: Bell’s Inequalities for Classical Conditional

Probabilities

Just as in the previous section suppose that we are given real numbers pi and pij in
[0, 1] with i = 1 . . . n and (i, j ) ∈ S. But now we ask: when can these numbers be
conditional probabilities of certain events and their conjunctions? More precisely:
given the numbers pi and pij , do there exist events Ai and ai (i = 1 . . . n) in a
classical probability space (, , p) such that pi and pij are the following classical
conditional probabilities:

pi = p(Ai |ai )
pij = p(Ai ∧ Aj |ai ∧ aj )

Or again in brief, do the numbers pi and pij admit a conditional Kolmogorovian

representation?
The answer here is more permissive. Except for some extremal values the
numbers pi and pij always admit a conditional Kolmogorovian representation: any
correlation vector p admits a conditional Kolmogorovian representation if p has no
(i, j ) ∈ S such that either (i) pi = 0 or pj = 0 but pij = 0; or (ii) pi = pj = 1 but
pij = 1.
Obviously, any Kolmogorovian representations is a conditional Kolmogorovian
representation with ai =  for all i = 1 . . . n. However, correlation vectors
admitting a conditional Kolmogorovian representation are not necessarily in the
classical correlation polytope.3

3 For example consider the following correlation vector in R(n, S) with n = 2 and S = {1, 2}:
 
2 2 1
p = , ;
3 3 5

The vector p violates Bell’s inequality

p1 + p2 − p12  1

hence it is not in c(n, S) and, consequently, it does not admit a Kolmogorovian representation.
However, p admits a conditional Kolmogorovian representation with the following atomic events
and probabilities:

1
p(A1 ∧ A2 ∧ a1 ∧ a2 ) =
25
4
p(A⊥ ⊥
1 ∧ A2 ∧ a1 ∧ a2 ) =
25
9
p(A⊥ ⊥ ⊥ ⊥
1 ∧ A2 ∧ a1 ∧ a2 ) = p(A1 ∧ A2 ∧ a1 ∧ a2 ) =
25
1
p(A⊥ ⊥ ⊥ ⊥ ⊥ ⊥
1 ∧ A2 ∧ a1 ∧ a2 ) = p(A1 ∧ A2 ∧ a1 ∧ a2 ) =
25
(The probability of all other atomic events is 0.)
362 G. Hofer-Szabó

Now, any correlation vector can be expressed as a convex combination of

(not necessarily classical) vertex vectors. A vertex vector u is defined as follows:
ui , uij ∈ {0, 1} for all i = 1 . . . n and (i, j ) ∈ S. Obviously, classical vertex vectors
are vertex vectors but not every vertex vector is classical. For example the vertex
vectors (0, 0; 1), (1, 0; 1), (0, 1; 1) or (1, 1; 0) in R(n, S) with n = 2 and S = {1, 2}
are not classical. There are 2n+|S| different vertex vectors u k in R(n, S) and 2n
different classical vertex vectors u ε .
Denote the convex hull of the vertex vectors by
⎧  ⎫
⎨  n+|S|
2 n+|S|
2 ⎬

u(n, S) := v ∈ R(n, S) v = λk u k ; λk  0; λk = 1
⎩  ⎭
k=1 k=1

Contrary to c(n, S), the polytope u(n, S) is not a simplex, hence the expansion of
the correlation vectors in u(n, S) by vertex vectors is typically not unique.4
Now, the set of correlation vectors admitting a conditional Kolmogorovian
representation is dense in u(n, S): all interior points of u(n, S) admit a conditional
Kolmogorovian representation and also all surface points, except for which there
is a pair (i, j ) ∈ S such that either (i) pi = 0 or pj = 0 but pij = 0; or
(ii) pi = pj = 1 but pij = 1. Denote the set of correlation vectors admitting a
conditional Kolmogorovian representation by u (n, S).
Now, let’s go over to the common causal explanation of conditional correlations.
Let p be a proper correlation vector in u (n, S). Lying in u (n, S) the correlation
vector p represents the conditional probabilities of certain events and their conjunc-
tions and the events associated to some pairs (i, j ) ∈ S are conditionally correlated:

p(Ai ∧ Aj |ai ∧ aj ) = p(Ai |ai ) p(Aj |aj ) (16.21)

Interpret now the events Ai and ai in the context of physical experiments: Let ai
represent a possible measurement that an experimenter can perform on an object.
Then the event ai ∧ aj will represent the joint performance of measurements ai
and aj . Let furthermore the event Ai represent an outcome of measurement ai

4 The correlation vector

 
2 2 1
p = , ;
3 3 5

for example can be expanded in many different ways:

1 1 1 2
p = (0, 1; 0) + (1, 0; 0) + (1, 1; 1) + (1, 1; 0)
3 3 5 15
1 1 7
= (0, 0; 0) + (1, 1; 1) + (1, 1; 0)
3 5 15
etc.
16 On the Three Types of Bell’s Inequalities 363

and Ai ∧ Aj represent an outcome of the jointly performed measurement ai ∧ aj .

Then (16.21) expresses a correlation between two measurement outcomes provided
their measurements have been performed.
When do the conditional correlations in a conditional Kolmogorovian represen-
tation of p have a common causal explanation?
The set of conditional correlations are said to have a non-conspiratorial joint
common causal explanation if there is a partition {Ck } (k ∈ K) of  such that for
any (i, j ) in S and k ∈ K

p(Ai ∧ Aj |ai ∧ aj ∧ Ck ) = p(Ai |ai ∧ Ck ) p(Aj |aj ∧ Ck ) (16.22)

p(ai ∧ aj ∧ Ck ) = p(ai ∧ aj ) p(Ck ) (16.23)

Equations (16.22) express the Reichenbachian idea that the common cause is to
screen off all correlations. Equations (16.23) express the so-called no-conspiracy,
the idea that common causes should be causally, and hence probabilistically,
independent of the measurement choices. The common causal explanation is joint
since all correlations (16.21) have the same common cause.
Now, suppose the correlations in the a given conditional Kolmogorovian repre-
sentation of (16.21) of p in u (n, S) have a non-conspiratorial joint common causal
explanation. Introduce again the notation

pik = p(Ai |ai ∧ Ck )

ck = p(Ck )


and consider the k common cause vectors of the correlation vector p:
 
p k = p1k , . . . , pnk ; . . . , pik pjk , . . .

We call a non-conspiratorial joint common cause deterministic if p k is a determin-

istic common cause vector for all k ∈ K; otherwise we call it indeterministic. We
call a deterministic non-conspiratorial joint common cause {Ck } (k ∈ K) a property;
and an indeterministic common cause a propensity.
Now, due to (16.22)–(16.23) and the theorem of total probability

p = ck p k
k

Since common cause vectors are independence vectors lying in c(n, S), their convex
combination also lies in the classical correlation polytope c(n, S). Thus, p being a
classical correlation vector is a necessary condition for a conditional Kolmogorovian
representation of p to have a non-conspiratorial joint common causal explanation.
364 G. Hofer-Szabó

Conversely, knowing the k common cause vectors and the probabilities ck alone,
one can construct the classical probability space  with the conditionally correlating
events Ai and ai and the common causes {Ck }.
Observe that the situation is now different from the one in the previous section:
the numbers pi and pij can be conditional probabilities of events and their
conjunctions even if they violate the corresponding Bell inequalities. However,
the conditional correlations between these events have a non-conspiratorial joint
common causal explanation if and only if the correlation vector composed of the
numbers pi and pij lies in the classical correlation polytope. In other words, in case
of classical conditional probabilities Bell’s inequalities do not test the conditional
Kolmogorovian representability but whether correlations can be given a common
causal explanation.

16.4 Relating Case 1 and Case 2

How do the Kolmogorovian and the conditional Kolmogorovian representation

relate to one another?
In this section I will show that (i) a conditional Kolmogorovian representation of
a classical correlation vector has a property explanation only if the representation
is non-signaling (see below); and (ii) a correlation vector p has a Kolmogorovian
representation if and only if it has a property explanation for any non-signaling
conditional Kolmogorovian representation.
Let p be a proper correlation vector in c(n, S) and consider a conditional
Kolmogorovian representation of p.  Obviously, there are many such representations
of p depending on the measurement conditions ai . Let’s say, somewhat loosely, that
a conditional Kolmogorovian representation has a property/propensity explanation
if the conditional correlations in the representation have a property/propensity
explanation.
First, we claim that a conditional Kolmogorovian representation of a correlation
vector p in c(n, S) has a property explanation only if the representation satisfies
non-signaling:

p(Ai |ai ) = p(Ai |ai ∧ aj ) (16.24)

p(Aj |aj ) = p(Aj |ai ∧ aj ) (16.25)

for any (i, j ) ∈ S. Note that non-signaling is not a feature of the correlation vector
itself but of the representation. For the same correlation vector in c(n, S) one can
provide both non-signaling and also signaling representations.
Now, a conditional Kolmogorovian representation can have a property explana-
tion only if it satisfies non-signaling. Recall namely that for a property {Ck }:

p(Ai |ai ∧ Ck ) ∈ {0, 1}

16 On the Three Types of Bell’s Inequalities 365

and hence

p(Ai |ai ∧ Ck ) = p(Ai |ai ∧ aj ∧ Ck ) (16.26)

for any i, j = 1 . . . n and k ∈ K. But (16.26) together with no-conspiracy (16.23)

and the theorem of total probability imply non-signaling (16.24)–(16.25). Thus,
satisfying non-signaling is a necessary condition for a conditional Kolmogorovian
representation to have a property explanation. Signaling conditional Kolmogorovian
representations do not have a property explanation.
Second, suppose that a conditional Kolmogorovian representation of p have a
property explanation. That is p has a conditional Kolmogorovian representation
in a classical probability space  and all the conditionally correlating event pairs
have a non-conspiratorial deterministic joint common cause in . Then these
properties (deterministic common causes) provide a Kolmogorovian (unconditional)
representation for p.
Observe namely that
 
p(Ai ∧ ai ) p(Ai ∧ ai ∧ Ck ) p(Ai |ai ∧ Ck )p(ai ∧ Ck )
pi = p(Ai |ai )= = k
= k
p(ai ) p(ai ) p(ai )

∗ p(Ai |ai ∧ Ck )p(ai )p(Ck )  
= k = p(Ai |ai ∧ Ck )p(Ck ) = p(Ck )
p(ai )
k k k: pi =1

∗ 
where the equation = holds due to no-conspiracy (16.23) and the symbol k: pik =1
means that we sum up for all k for which pik = 1. Similarly, using (16.22)–(16.23)
one obtains

pij = p(Ck )
k: pik =1, pjk =1

That is the events

Ci = ∨k: pk =1 Ck
i

Ci ∧ Cj = ∨k: pk =1, pk =1 Ck
i j

provide a Kolmogorovian representation for the numbers pi and pij .

Third, conversely, if p admits a Kolmogorovian representation, then it also
admits a property explanation for any non-signaling conditional Kolmogorovian
representation.
Namely, if p admits a Kolmogorovian representation, then there is a partition
{Cε } in  such that
366 G. Hofer-Szabó

 
pi = p(Ci ) = p(Cε ) = λε (16.27)
ε: ε i =i ε: ε i =i
 
pij = p(Ci ∧ Cj ) = p(Cε ) = λε (16.28)
ε: ε i =1, ε j =1 ε: ε i =1, ε j =1

 that
Now, consider a non-signaling conditional Kolmogorovian representation of p,
is let there be events Ai and ai in   such that

pi = p(Ai |ai )
pij = p(Ai ∧ Aj |ai ∧ aj )

and suppose that non-signaling (16.24)–(16.25) hold for any (i, j ) ∈ S.

Now, the events Cε provide a propensity explanation for the conditional Kol-
mogorovian representation of p in the following sense. First, let ε, ε  ∈ {0, 1}n and
define the events aε in   as follows:

ai if ε i = 1
aε := ∧i aεi where aεi =
a i if ε i = 0

Then, extend the algebra   to   generated by the atomic events Dε,ε defined as
follows:

∨ε Dε,ε = aε (16.29)

∨ε Dε,ε = Cε (16.30)
p(Dε,ε ) = p(aε ) p(Cε ) = p(aε ) λε (16.31)
p(Ai |Dε,ε ) = εi piε (16.32)
p(Ai ∧ Aj |Dε,ε ) = εi piε εj pjε (16.33)

where the numbers piε ∈ [0, 1] will be specified below. Now, using (16.29)–(16.33)
one obtains

p(Ai ∧ Aj |ai ∧ aj ∧ Cε ) = piε pjε = p(Ai |ai ∧ Cε ) p(Aj |aj ∧ Cε )


p(ai ∧ aj ∧ Cε ) = p(Dε,ε ) = p(ai ∧ aj ) p(Cε )
ε : ε i =1, ε j =1

That is the partition {Cε } provides a propensity explanation for the conditional
Kolmogorovian representation of p with probabilities specified in (16.27)–(16.28).
Now
16 On the Three Types of Bell’s Inequalities 367

 
p(Ai ∧ aε ) p(Ai ∧ Dε,ε ) p(Ai |Dε,ε ) p(Dε,ε )
p(Ai |aε ) = = ε
= ε
p(aε ) p(aε ) p(aε )
  ε
ε p p(aε )λε 
= ε i i = ε i piε λε
p(aε ) ε

that is

pi = p(Ai |ai ) = piε λε (16.34)
ε

and similarly

pij = p(Ai ∧ Aj |ai ∧ aj ) = piε pjε λε (16.35)
ε

Composing 2n independence vectors p ε from the numbers piε :

 
p ε = . . . piε . . . pjε . . . ; . . . piε pjε . . .

(16.34)–(16.35) reads as follows:


p = λε p ε
ε

This means that the numbers piε are to be taken from [0, 1] such that the
2n independence vectors p ε provide a convex combination for p with the same
coefficients λε as the vertex vectors u ε do. One such expansion always exists.
Namely, when p ε = u ε . In this case piε = εi for any i = 1 . . . n and ε ∈ {0, 1}n
and (16.34)–(16.35) reads as follows:

pi = p(Ai |ai ) = εi p(Cε ) = p(Ci )
ε

pij = p(Ai ∧ Aj |ai ∧ aj ) = ε i εj p(Cε ) = p(Ci ∧ Cj )
ε

That is we obtain a property explanation for the conditional Kolmogorovian

representation of p.  However, for classical correlation vectors which are not
classical vertex vectors one can also provide for p various convex combinations
by indeterministic independence vectors with the coefficients λε (see (16.18) as an
example). In this case we obtain a propensity explanation for the given conditional
Kolmogorovian representation of p. 
To sum up, a correlation vector p has a Kolmogorovian representation if and
only if it has a property explanation for any non-signaling conditional Kolmogoro-
vian representation. This equivalence justifies retrospectively why we used the
368 G. Hofer-Szabó

term “property” both for a deterministic common cause in the conditional Kol-
mogorovian representation and also as a synonym of “event” in the unconditional
Kolmogorovian representation.

16.5 Case 3: Bell’s Inequalities for Quantum Probabilities

Again start with real numbers pi and pij in [0, 1] with i = 1 . . . n and (i, j ) ∈ S.
But now the question is the following: when can these numbers be quantum
probabilities? That is, given the numbers pi and pij , do there exist projections
Ai (i = 1 . . . n) representing quantum events in a quantum probability space
(P(H ), ρ), where P(H ) is the projection lattice of a Hilbert space H and ρ is a
density operator on H representing the quantum state, such that pi and pij are the
following quantum probabilities:

pi = Tr(ρAi )
pij = Tr(ρ(Ai ∧ Aj ))

(Here Ai ∧ Aj denotes the projection projecting on the intersection of the closed

subspaces of H onto which Ai and Aj are projecting.) Again in brief, do the
numbers pi and pij admit a quantum representation?
The answer is again given by Pitowsky (1989, p. 72). Introduce the notion of a
quantum vertex vector. A quantum vertex vector is a vertex vector in R(n, S) such
that

uiε = εi i = 1...n
uijε  εi ε j (i, j ) ∈ S

Obvious, any classical vertex vector is a quantum vertex vector and any quantum
vertex vector is a vertex vector, but the reverse inclusion does not hold. In R(n, S)
with n = 2 and S = {1, 2} for example the vertex vector (1, 1; 0) is quantum but not
classical, and the vertex vectors (0, 0; 1), (1, 0; 1), (0, 1; 1) are not even quantum.
Denote by q(n, S) the convex hull of quantum vertex vectors. Pitowsky then
shows that almost all correlation vectors in q(n, S) admit a quantum representation.
More precisely, Pitowsky shows that—denoting by q  (n, S) the set of quantum
correlation vectors that is the set of those correlation vectors which admit a quantum
representation—the following holds:
(i) c(n, S) ⊂ q  (n, S) ⊂ q(n, S);
(ii) q  (n, S) is convex (but not closed);
(iii) q  (n, S) contains the interior of q(n, S)
We can add to this our result in the previous section:
(iv) q(n, S) ⊂ u (n, S) ⊂ u(n, S)
16 On the Three Types of Bell’s Inequalities 369

Thus, the set of numbers admitting a quantum representation is strictly larger than
the set of numbers admitting a Kolmogorovian representation but strictly smaller
than the set of numbers admitting a conditional Kolmogorovian representation.
Let us turn now to the question of the common causal explanation. Let p be a
proper quantum correlation vector. p then represents the quantum probabilities of
certain quantum events and their conjunctions and the events associated to some
pairs (i, j ) ∈ S are correlated:

Tr(ρ(Ai ∧ Aj )) = Tr(ρAi ) Tr(ρAj ) (16.36)

When do the correlations (16.36) have a common causal explanation?

The set of quantum correlations has a joint quantum common causal explanation
if there is a partition {Ck } (k ∈ K) in P(H ) (that is a set of mutually orthogonal
projection adding up to the unity) such that for any (i, j ) in S and k ∈ K

Tr(ρ k (Ai ∧ Aj )) = Tr(ρ k Ai ) Tr(ρ k Aj ) (16.37)

where
Ck ρ Ck
ρ k :=
Tr(ρ Ck )

is the density operator ρ after a selective measurement by Ck . If the partition {Ck } is

commuting with the each correlating pair (Ai , Aj ), then we call the common causes
commuting, otherwise noncommuting.
Now, suppose the correlations (16.36) have a joint quantum common causal
explanation. Does it follow that p is in c(n, S) that is Bell’s inequalities are satisfied?
Introduce again the notation

pik = Tr(ρ k Ai )
ck = Tr(ρ Ck )


and consider the k common cause vectors of the correlation vector p:
 
p k = p1k , . . . , pnk ; . . . , pik pjk , . . .

The common cause vectors are independence vectors. Hence, using (16.37) and the
theorem of total probability, the convex combination of the common cause vectors

p c = ck p k (16.38)
k

will be in c(n, S). However p c is not necessarily identical with the original
 More precisely, p c = p for any ρ if {Ck } are commuting
correlation vector p.
370 G. Hofer-Szabó

common causes. But if {Ck } are noncommuting common causes, then p c and p
can be different and hence even if p c ∈ c(n, S), p might be outside c(n, S).
In short, a quantum correlation vector can have a joint noncommuting common
causal explanation even if it lies outside the classical correlation polytope. The
correlation vector p is confined in c(n, S) only if the common causes are required
to be commuting.5
To sum up, in the case of quantum probabilities we found a scenario different from
both previous cases. Here Bell’s inequalities neither put a constraint on whether
numbers can be quantum probabilities nor on whether the correlation between
events with the prescribed probability can have a common causal explanation.
Bell’s inequalities constrain common causal explanations only if common causes
are understood as commuting common causes.
Perhaps it is worth mentioning that in algebraic quantum field theory (Rédei
and Summers 2007; Hofer-Szabó and Vecsernyés 2013) and quantum information
theory (Bengtson and Zyczkowski 2006) the violation of Bell’s inequalities com-
posed of quantum probabilities is used for another purpose: it places a bound on
the strength of correlations between certain events. Abstractly, one starts with two
mutually commuting C ∗ -subalgebras A and B of a C ∗ -algebra C and defines a Bell
operator R for the pair (A, B) as an element of the following set:

1 
B(A, B) := A1 (B1 + B2 ) + A2 (B1 − B2 )  Ai = A∗i ∈ A;
2
(
∗
Bi = Bi ∈ B; −1  Ai , Bi  1

where 1 is √the unit element of C. Then one can prove that for any Bell operator R,
|φ(R)|  2 for any state φ; but |φ(R)|  1 for separable states (i.e. for convex
combinations of product states).
In other words, in these disciplines one fixes a Bell operator (a witness operator)
and scrolls over the different quantum states to see which of them is separable.
In this reading, Bell’s inequalities neither test Kolmogorovian representation nor
common causal explanation but separability of certain normalized positive linear
functionals. Obviously, this role of Bell’s inequalities is completely different from
the one analyzed in this paper.

16.6 The EPR-Bohm Scenario

Bell’s inequalities in the EPR-Bohm scenario are standardly said to be violated. But
which Bell inequalities and what does this violation mean?

5 For the details and for a concrete example see Hofer-Szabó and Vecsernyés (2013, 2018).
16 On the Three Types of Bell’s Inequalities 371

Let us start in the reverse order, with quantum representation. Consider the
EPR-Bohm scenario with pairs of spin- 12 particles prepared in the singlet state. In
quantum mechanics the state of the system and the quantum events are represented
by matrices on M2 ⊗ M2 , where M2 is the algebra of the two-dimensional complex
matrices. The singlet state is represented as:

1  3

ρs = 1⊗1− σk ⊗ σk
4
k=1

and the event that the spin of the particle is “up” on the left wing in direction ai
(i = 1, 2); on the right wing in direction bj (i = 3, 4); and on both wings in
directions ai and bj , respectively, are represented as:

1 
Ai = (1 + ai · σ ) ⊗ 1
2
1 
Aj = 1 ⊗ (1 + bj · σ )
2
1 
Aij = (1 + ai · σ ) ⊗ (1 + bj · σ )
4

where 1 is the identity matrix on M2 , σ = (σ 1 , σ 2 , σ 3 ) is the Pauli vector, and ai

(i = 1, 2) and bj (j = 3, 4) are the spin measurement directions on the left and
right wing, respectively. Furthermore, Aij = Ai Aj = Ai ∧ Aj since Ai and Aj are
commuting.
The quantum probabilities are generated by the trace formula:

1
pi = Tr(ρ s Ai ) = (16.39)
2
1
pj = Tr(ρ s Aj ) = (16.40)
2
 
1 θ ij
pij = Tr(ρ s Aij ) = sin2 (16.41)
2 2

where θ ij denotes the angle between directions ai and bj . As it is well known, for
the measurement directions
1
a1 = (0, 1, 0) b3 = √ (1, 1, 0)
2
1
a2 = (1, 0, 0) b4 = √ (−1, 1, 0)
2

the Clauser-Horne inequality

372 G. Hofer-Szabó

√
1+ 2
− 1  p13 + p23 + p14 − p24 − p1 − p3 = − (16.42)
2
is violated.
What does the violation of the Clauser-Horne inequality (16.42) mean with
respect to the quantum representation? As clarified in Sect. 16.5, it does not
mean that the numbers (16.39)–(16.41) cannot be given a quantum mechanical
representation. Equations (16.39)–(16.41) just provide one. On the other hand, the
violation of (16.42) neither means that the EPR correlations

pij = Tr(ρ s Aij ) = Tr(ρ s Ai )Tr(ρ s Aj ) = pi pj i, j = 1, 2 (16.43)

cannot be given a joint quantum common causal explanation. In Hofer-Szabó and

Vecsernyés (2012, 2018) we have provided a partition {Ck } such that the screening-
off conditions (16.37) hold for all EPR correlations (16.43). That is {Ck } is a joint
common causal explanation for all the four EPR correlations.6 As also shown in
Sect. 16.5, these joint common causes need to be noncommuting common causes,
since commuting common causes would imply the Clauser-Horne inequality (16.42)
which is violated in the EPR-Bohm scenario. Thus, the violation of the Clauser-
Horne inequality (16.42) does not exclude a common causal explanation of the
EPR-Bohm scenario—as long as noncommuting common causes are tolerated in
the explanation.7
To see the thrust of the violation of the Clauser-Horne inequalities, we have to go
over to the interpretations of quantum mechanics. As for the operational interpreta-
tion, the violation of the Clauser-Horne inequalities, as clarified in Sect. 16.3, again
does not mean that the numbers (16.39)–(16.41) cannot be given an operational
interpretation. There is such an interpretation; otherwise quantum mechanics would
not be empirically adequate. The operational interpretation is the following:

pi = p(Ai |ai )
pj = p(Bj |bj )
pij = p(Ai ∧ Bj |ai ∧ bj )

where ai denotes the event that the spin on the left particle is measured in direction
ai , and Ai denotes the event that the outcome in this measurement is “up”.
(Similarly, for bj and Bj .) That is, the probabilities are conditional probabilities of
certain measurement outcomes provided that certain measurements are performed.
However, the violation of the Clauser-Horne inequality (16.42) does exclude that
the conditionally correlating pairs of outcomes

6 More than that, we have also shown that in an algebraic quantum field theoretic setting the
common causes can even be localized in the common past of the correlating events.
7 The problem with such noncommuting common causes, however, is that it is far from being clear

how they should be interpreted.

16 On the Three Types of Bell’s Inequalities 373

pij = p(Ai ∧ Aj |ai ∧ aj ) = p(Ai |ai ) p(Aj |aj ) = pi pj i, j = 1, 2

has a non-conspiratorial joint common causal explanation. Thus, in the operational

interpretation, Bell’s inequalities filter common causal explanations.
Finally, let us turn to the ontological interpretation. As is clarified by Pitowsky,
the violation of the Clauser-Horne inequality (16.42) excludes the numbers (16.39)–
(16.41) to be classical unconditional probability of certain events or properties and
their conjunctions. That is, there are no such numbers Ai in a classical probability
space such that

pi = p(Ai )
pj = p(Bj )
pij = p(Ai ∧ Bj )

Consequently, there are no correlations

pij = p(Ai ∧ Aj ) = p(Ai ) p(Aj ) = pi pj i, j = 1, 2

and a fortiori no need to look for a common causal explanation.

As shown in Sect. 16.4, a correlation vector has a Kolmogorovian representation
if and only if it has a property explanation for any non-signaling conditional
Kolmogorovian representation. The probabilities in the EPR-Bohm scenario are
non-signaling:

pi = p(Ai |ai ) = p(Ai |ai ∧ bj )

pj = p(Ai |bj ) = p(Bj |ai ∧ bj )

Hence, the violation of (16.42) again excludes a property explanation for any non-
signaling conditional Kolmogorovian representation.

16.7 Conclusions

What does it mean that a set of numbers violates Bell’s inequalities? One can answer
this question in three different ways depending on whether the numbers are inter-
preted as classical unconditional probabilities, classical conditional probabilities, or
quantum probabilities:
(i) In the first case, the violation of Bell’s inequalities excludes these numbers
to be interpreted as the classical (unconditional) probabilities of certain
events/properties and their conjunctions. The satisfaction of Bell’s inequalities
374 G. Hofer-Szabó

does not only guarantee the existence of such events, but also the existence
of a set of joint common causes screening off the correlations between these
events.
(ii) In the second case, the violation of Bell’s inequalities does not exclude that
these numbers can be interpreted as the classical conditional probability of
certain events (measurement outcomes) conditioned on other events (measure-
ment settings). However, it does exclude a non-conspiratorial joint common
causal explanation for the conditional correlations between these events.
(iii) Finally, the violation or satisfaction of Bell’s inequalities has no bearing either
on whether a set of numbers can be interpreted as quantum probability of
certain projections, or whether there can be given a joint common causal
explanation for the correlating projections—as long as noncommuting com-
mon causes are adopted in the explanation.

Acknowledgements This work has been supported by the Hungarian Scientific Research Fund,
OTKA K-115593 and a Research Fellowship of the Institute of Advanced Studies Köszeg. I wish
to thank Márton Gömöri and Balázs Gyenis for reading and commenting on the manuscript.

References

Bengtson, I., & Zyczkowski, K. (2006). Geometry of quantum states: An introduction to quantum
entanglement. Cambridge: Cambridge University Press.
Gömöri, M., & Placek, T. (2017). Small probability space formulation of Bell’s theorem. In G.
Hofer-Szabó & L. Wronski (Eds.), Making it formally explicit – Probability, causality and
indeterminism (European studies in the philosophy of science series, pp. 109–126). Dordrecht:
Springer Verlag.
Hofer-Szabó, G., & Vecsernyés, P. (2012). Noncommuting local common causes for correlations
violating the Clauser–Horne inequality. Journal of Mathematical Physics, 53, 12230.
Hofer-Szabó, G., & Vecsernyés, P. (2013). Bell inequality and common causal explanation in
algebraic quantum field theory. Studies in History and Philosophy of Science Part B – Studies
in History and Philosophy of Modern Physics, 44(4), 404–416.
Hofer-Szabó, G., & Vecsernyés, P. (2018). Quantum theory and local causality. Dordrecht:
Springer Brief.
Hofer-Szabó, G., Rédei, M., & Szabó, L. E. (1999). On Reichenbach’s common cause principle and
on Reichenbach’s notion of common cause. The British Journal for the Philosophy of Science,
50, 377–399.
Hofer-Szabó, G., Rédei, M., & Szabó, L. E. (2013). The principle of the common cause.
Cambridge: Cambridge University Press.
Pitowsky, I. (1989). Quantum probability – Quantum logic. Dordrecht: Springer.
Rédei, M., & Summers, J. S. (2007). Quantum probability theory. Studies in History and
Philosophy of Modern Physics, 38, 390–417.
Reichenbach, H. (1956). The direction of time. Los Angeles: University of California Press.
Szabó, L. E. (2008). The Einstein-Podolsky-Rosen argument and the Bell inequalities. Internet
encyclopedia of philosophy, http://www.iep.utm.edu/epr/.
Chapter 17
On the Descriptive Power of Probability
Logic

Ehud Hrushovski

Abstract In a series of imaginary conversations, a philosopher of physics and

a mathematical logician discuss the investigation of a mathematical relational
universe and the description of the world by physical theory, as metaphors for
each other. They take for granted that basic events can be observed directly, as
atomic formulas, and ask about the ability to interpret deeper underlying structures.
This depends on the logic used. They are interested specifically in the role of
classical probability logic in this endeavor. Within the model they adopt, as a case
study, probability logic is highly competent in discovering and describing spatial
structures; where a very general notion of space is taken to correspond to the logical
notion of monadic, both with natural relativistic corrections. However, this logic
can go no further in itself, quite unlike the unbounded abilities of first-order logic.
In some cases new first-order structures emerge at the point of failure of probability
logic.

Keywords Probability logic · Model theory · Expressive power

17.1 Introduction: Logic as a Descriptive Tool

Logic is often defined as the “systematic study of the form of valid inference”.1 “Let
C be the planets, B to be near, A not to twinkle,” wrote Aristotle; you can then study
the implication from the premises (∀x)(A(x) → B(x)) and (∀x)(C(x) → A(x)),
in modern notation, to the conclusion that the planets are near. It is the pride and
weakness of his method that “no further term is required from without in order to

1 From the wikipedia entry on Logic, retrieved June 2019.

E. Hrushovski ()
Mathematical Institute, Oxford, UK

© Springer Nature Switzerland AG 2020 375

M. Hemmo, O. Shenker (eds.), Quantum, Probability, Logic,
Jerusalem Studies in Philosophy and History of Science,
https://doi.org/10.1007/978-3-030-34316-3_17
376 E. Hrushovski

make the consequence necessary” (Aristotle 1906). By the end of the argument, we
may learn a new fact about B and C; we will not be introduced to any new idea D.
Modern logic has, it seems to me, an additional role: the introduction of new
concepts. The primary use of the first-order relational calculus is not for judging the
wisdom of asserting a certain hypothesis, whose meaning is assumed clear. It is first
of all a tool for carving out new ideas and dividing lines, about which assertions
can subsequently be formulated. An ancient example is the notion of parallel lines,
defined from a basic relation of incidence between a point and a line. With the aid of
a quantifier, concerning the (non-) existence of an intersection point, a new relation
– parallelism – has been defined from one taken to be more basic (a point lying on
a line). Taking such definitions further, Descartes shows how to interpret algebraic
operations in geometry, in such a way that one can turn back and define the geometry
in the algebra. Or to move to the nineteenth century, take the definition of continuity
of a function F :

(∀x)(∀ > 0)(∃δ > 0)(∀y)(d(x, y) < δ → d(F (x), F (y)) < )

Continuity is clearly a new concept; yet it somehow emerges from a background

consisting only of an idea of distance d, with ordered values, using the first-order
logical symbols alone.
To be sure, a concept like continuity does not reduce to the formal definition; it
continues to have associated pictures and ideas, without which we would drown
in an exponentially growing set of potential formal proofs. But if mathematics
as a human activity would make little sense without the conceptual imagery, the
specifically mathematical nature of this activity does have much to do with the
casting of these images in the clay of logic. Having proved itself for a certain time,
the formal definition becomes the final arbiter; more than that, it informs and guides
the future evolution of the initial idea. In any case, whatever one’s position on this,
at the very least we have here a formal avatar of concept creation.
It is the dynamic of alternation of quantifiers that has this dramatic creative
power; it encodes an activity: given x and , search for a good δ, whose discovery
may lead to questions about a new x requiring a new witness. Alternation is only
possible if the relations we begin with are at least binary, and thus have n-ary
boolean combinations for any n. The formula

(∀y)(d(x, y) < δ → d(F (x), F (y)) < )

relates 3 variables, x, , δ, in a new way, and we can now afford to quantify one
out. By contrast, if we only have unary relations, quantifiers produce nothing new
beyond sentences; there the monadic expressive power ends. We obtain assertions
whose truth we can try to decide, but no new definitions, no new distinctions among
elements populating the given world. This is the fundamental difference between
Aristotle’s logic, and the modern version.
It is easy to multiply examples of the birth of new concepts in first-order
language; mathematics forms a vast web of fields of discourse, whose basic concepts
17 On the Descriptive Power of Probability Logic 377

are introduced and inter-related in this manner. Sets are sometimes used, but set
theory itself is understood via first-order logic; the primary use of the assumption
that Y consists of all subsets of X is the first-order comprehension axiom.
The interpretation of mathematics in set theory does show, in one way among
many, that one binary relation – the smallest non-monadic relational language –
suffices to interpret anything we like. And having properly started, Gödel’s theorem
tells us it would be impossible to set fixed limits to this process.
The mathematics used in physics is no exception. Science is only at the very
end (if that) about checking the truth of universal sentences. There is first the step
of recognizing or creating the notions, in terms of which the theory is formulated
and may yield universal consequences. Regular viewings of a light point in the
night sky are interpreted in terms of a celestial body; an orbital ellipse describes its
movement; a differential equation predicts the ellipse. Newton’s laws are phrased
in terms of derivatives; when fully clarified, they are formulated in terms of tangent
and cotangent spaces, vector fields and flows and symplectic forms. The definition
of all of these, and their connection to the simpler mathematical objects that have
a direct interpretation – the distance a stone will fly – is carried out using the first-
order quantifiers.
It is in this light that I would like to consider probability logic, where the
absolute quantifiers are replaced by numerical ones giving probabilities.2 In the
setting of Mendelian genetics, say, an ordinary quantifier may be used to ask
whether a yellow pea ever descends from two orange ones. A probability quantifier
rather asks: what proportion of such descendents is yellow? If we pick one at
random, what is the likelihood that it is yellow? Robust logical calculi where
such expressions are permitted have been constructed.3 Such systems admit a
completeness theorem, so that the (false) dichotomy between probabilistic and
measure-theoretic language becomes a useful duality, with the measure theory on
the semantic side complementing the random-variable viewpoint in the language.
They are applicable when a natural measure forms part of the structure of the
universe we are trying to describe.
In certain circumstances, probability quantifiers significantly enrich the power of
the logic. A trivial way this may happen is at the very reading of the initial data. A
graph on a large number n of points, obtained by a random choice of edges, looks
the same to first-order eyes regardless of whether the edge probability is 1%, or
99%, provided n is big enough compared to the formulas considered. Probability
quantifiers, by their design, will of course directly and immediately detect the
difference.
But is probability logic of use beyond the simple access to the basic data? Does it
also help in the formulation of new concepts, new relations, and if so, is it possible

2 SeeMumford (2000) for a contrasting view.

3 We will work with a finitary first-order calculus close to that of Keisler (1985) Sect. 4.1; the
importance of a theory of approximations and limiting behavior favors this over the infinitary
versions.
378 E. Hrushovski

to say in what way? This is the question we attempt to answer here. This is quite
different from the obvious importance of probability or measure as mathematical
tools in their own right, studied from the outside using ordinary quantifiers; we are
interested in the creative power coded within that calculus.
Consider this thought-experiment. You are plunged into a world of which nothing
is known; or perhaps you sit outside, but have the possibility of probing it. It is
a Hilbert-style world, of the general kind of his axiomatization of geometry; with
recognizable ‘things’ (perhaps points and lines, ‘material particles’ (Maxwell 1925),
planets, packets of radiation, small colored pixels, · · · . ) Some basic relations among
them – we will start with unary and binary relations – can be directly sensed. If
you are unlucky, your world consists of Kolmogorovian noise; no pattern can be
inferred, no means of deducing any additional observations from some given ones.
But if there is a pattern, how much of it you will discern depends not only on your
eyes, but on the logical language you have available. Quite possibly, the laws of the
world in question can only be formulated in terms of concepts and structures that
are not immediately given in the basic relations. If you use first-order logic, starting
with what seem to be points and lines, interpreted in algebra and leading to number
theory, such entities may be discovered and analyzed; there is no limit to the level
of abstraction that may be needed. Is there a similar phenomenon if your logic is
purely stochastic? What features will you be able to make out?
This essay is conceived as a dialogue between a mathematical logician (L), and
a philosopher of physics (P), sometimes joined by P’s brilliant student (S). The
philosopher, over L’s objections that he is engaged in purely technical lemmas, asks
L to present his thoughts on the expressive power of probability logic. This leads
to a number of conversations, in which the interlocutors attempt a philosophical
interpretation of the lemmas in question, and their mathematical background.
To try to gain insight into this question, they try to study a purified extract; a logic
with probability quantifiers replacing the usual ones entirely.
Their first conclusion is that the probability logic considered excels in char-
acterizing, and recognizing, space-like features. If the given universe includes a
metric giving a notion of distance, the metric space may be characterized by the
stochastic logical theory (a theorem of Gromov.) Under certain finite-dimensionality
assumptions, the logic is sometimes able to distinguish infinitesimal features, even
from coarse-grained data: if the metric is not directly given at the smallest scale, but
only some mid-scale relation depending on it is directly observable, nevertheless the
probability quantifiers can penetrate into the small scale structure of the world. If
we further assume that the universe (with respect to certain relations) has the same
statistical properties at every point, it can be shown to approximate a Riemannian
homogeneous space, a rather unique kind of mathematical object.
But there it ends. Probability logic can describe a locally compact space-like
substrate of a given relational world; and color this space by the local correlation
functions of some conjunctions of the basic relations. Beyond this, it can go no
further on its own. Conditioned on this spatial structure, all (or almost all) events
appear statistically independent; hidden structures beyond the monadic space may
exist, but can not be detected with stochastic logic.
17 On the Descriptive Power of Probability Logic 379

In some cases, hints of a deeper structure can be gleaned from the probability-
zero sets where the statistical independence breaks down; they must be analyzed by
other means. The general formulation of concepts in probability logic, with both
ordinary and probability quantifiers present, can be viewed then as an iteration
of a first-order layer, then a purely probabilistic one, and so on. The role of the
probability quantifiers appears restricted to a determination of the space associated
with the current conceptual framework, along with some correlations along it;
certain measure zero sets suggest new phenomena to be explored, once again with
first order logic.
Does the picture change if we use ternary or higher relations? Here a full
mathematical description is not yet available; but it is clear that the expressive power
of pure probability logic remains limited, and perhaps again restricted to specifying
parameters for a somewhat more generalized notion of space; the new phenomena
that occur will be briefly considered.
This essay is dedicated to the memory of my friend Itamar Pitowsky. The first
lines of the dialogue really do represent a meeting with Itamar, in the academic
year 2009/2010, on the Givat Ram campus. I was already thinking about related
questions, and Itamar was – as always – interested. But we first diverged to different
subjects, and the rest could never materialize in reality; these dialogues are purely
imaginary.
Needless to say, P does not stand for Pitowsky; I could not possibly, and did not
try, to guess Itamar’s real reactions.
One of the points that arises is the distinction between monadic logic and full
first order logic, and the apparent non-existence of any intermediate (beyond the
relativistic correction that is discussed in the text.) There is at least an analogy,
probably a deeper connection, between this and the question of existence of a
mathematical world lying strictly between linear algebra (linear equations) and full
algebraic geometry (all algebraic equations.) In this Itamar was interested over many
years, for reasons stemming (in part) from his deep study of Descarte’s geometry,
and the significance of quadratic equations there.4 I was likewise interested for
quite different reasons – such a world flickers briefly, intriguingly, in and out
of existence within Zilber’s proof5 (by contradiction) of the linearity of totally
categorical structures. I am still not aware of such a world, or a framework in which
it may be formulated.‘

17.2 Continuous Logic

P. L! You are back. What are you thinking about these days?
L. Actually, it is related to probability logic; but purely classical, Kolmogorov-
style probability, I’m afraid. A certain statement I can’t quite put my hands on;
perhaps it will end up requiring some higher-categorical language.

4 See Pitowsky (1999).

5 Zilber (1984), see also Pillay (1996).
380 E. Hrushovski

P. I’d be very interested to hear.

L. It’s a bit technical, I’d need to explain first a specific implementation of the
logic, and preliminary to that, the continuous logic it is embedded in.
P. All the better! Let it be several strolls. But here is my student Sagreda, perhaps
she will join us?
L. Let us start with continuous logic, then. Probabilities, like distances or masses,
are real numbers. We could discuss possible probabilities one by one, with relation
symbols Rα asserting: the probability of P is α. But that would be clumsy; it
would not capture the basic feature of continuity of probabilities. We think of
probability 0.5 as just barely distinguishable from 0.50000001; but if we set them
up discretely, they appear to be just as different as 0 and 1. It is preferable to use
a logic whose values admit a compatible notion of approximation. In real-valued
continuous logic,6 a basic relation symbol R comes with a possible range of values
– say, a closed interval on the real line. To specify a structure, for each assignment
of elements to the variables, we need to give a value for R. In two-valued logic, a
typical predicate has two possible outcomes, in a given possible world: Socrates is
mortal, or immortal; the mortality of Socrates is a fact, or a falsehood. In real-valued
logic, such two-valued predicates are still allowed. But so are terms such as “height”,
or “mass” or “temperature”; in these cases the possible values are real numbers. We
will take them to lie in some bounded range, known in advance; though ‘unbounded’
modifications of the logic also exist.7 This is not the ‘fuzzy logic’ viewpoint; one
should not think of the values as measures of the degree of truth of a predicate, but
rather that one asks questions whose answers are not expected to be binary in the
first place.
P. But in discrete first order logic, we allow in any case infinitely many relations.
If we wish to have a relation R taking values between 0 to 1, say, couldn’t we
equally well dedicate infinitely many ordinary predicates Rn for the purpose, with
Rn (x) specifying the n’th binary digit of R(x)?
L. This is a very good way of understanding continuous logic formulas. The one
difference lies in allowing for connectedness of the real line. Continuous logic treats
a structure giving a reading of 0.99999· · · as identical to one where the reading is
1.0000 · · · .
This is not too important for a single structure. But many of the fundamental
notions of logic concern the approximation of one structure by another. Recall for
example that a substructure A of a structure B is called an elementary submodel if
an observer living within A has no way of sensing that it is not the entire universe B;
any formula φ(x) with parameters from A having a solution b in B, has a solution
b in A too.

6 SeeChang and Keisler (1966), Ben Yaacov et al. (2008), and Ben Yaacov and Usvyatsov (2010).
7 Ben Yaacov (2008). One way to present such structures is to introduce a metric and use it at both
ends, with 0 an ideal limit of increasing resolution, and ∞ an ideal upper limit of distances; so
formulas, when considered up to a given precision, can only access a bounded region, as well as
only distinguish points at some resolution bounded above 0.
17 On the Descriptive Power of Probability Logic 381

Now suppose A is an elementary submodel in the usual 2-valued sense, and let
us use your coding. If R(b) takes a real value α = 0.abc · · · and A is an elementary
submodel, then one can find b in A with value 0.abc · · · , agreeing with α for as
many digits as we wish. In particular, R(b ) is arbitrarily close to α; this is indeed
what we would want. But the fact that the specifically chosen binary expansions
agree is very slightly stronger, and not a natural demand.
P. I see. So the right notion of approximation in continuous logic is simply this:
in continuous logic, A is an elementary submodel of B if for any formula φ(x) with
parameters from A, if φ(b) has value r ∈ R, then for any  > 0 one can find b in A
with φ(b ) of value between r −  and r + .
S. I think I can guess what connectives must be. Returning to sequences of
formulas in ordinary logic, say R1 , R2 , · · · and S1 , S2 , · · · , we can produce a
sequence T1 , T2 , · · · where each Tn is a Boolean function of finitely many Ri and
Sj ; say T1 = R1 &S1 , T2 = R1 ∨ S3 , and so on. In terms of binary expansions of
numbers, we obtain any function C(x, y) such that the n’th binary digit of C(x, y)
depends only on the first m digits of x and of y, for some m. I suppose the continuous
logic analogue would be any function from R2 to R, whose value up to an -accuracy
depends only on the inputs up to some δ?
L. Just so; in other words, you are simply describing a continuous function.
P. Isn’t it excessive to allow any continuous function as a connective? True, in the
two-valued case, too, we have an exponential number of n-place connectives. But
we can choose a small basis, for instance conjunction and negation, that generate
the rest.
L. In the continuous logic case, one can also approximate any connectives with
arbitrary accuracy using a small basis, for instance just the binary connective giving
the minimum of two values, multiplication by rationals as unary operators, and
rational constants.
P. And what about quantifiers?
L. Here a natural basis consists of sup and inf; see Ben Yaacov and Usvyatsov
(2010). There is an equivalent, perhaps more philosophically satisfying approach in
Chang and Keisler (1966), where a quantifier is seen to be an operator from a set
of values to a single value, continuous in the appropriate sense. Thus the existential
quantifier is viewed as the operator telling us if a set includes 1 or not; (∃y)φ(y) is
true in M if 1 ∈ {[φ(a)] : a ∈ M}.
P. After describing the usual predicate calculus, one steps back into the metalan-
guage to define the notion of logical validity.
L. It looks the same here. One can ask about the set of possible truth values of a
sentence σ in models M; or for simplicity, just whether that set reduces to a single
element {α}. This will be the case if and only if for any given  > 0, a syntactic
proof can be found (in an appropriate system) that shows the value to be within 
of α.
In fact all the major theorems of first order logic persist, as if they were made
for this generalization. Compactness, effectiveness statements remain the same;
notions such as model completeness and criteria for quantifier elimination; stability
and beyond. It took substantial work of a number of model theorists over several
382 E. Hrushovski

generations to get us used to the intrusion of real numbers; but once the right point of
view was found, it became clear that continuous logic is a very gentle generalization
of the old.
P. We have so far paid no attention to equality.
L. The analogue of equality in continuous logic can be taken to be a metric, the
data of the distance between any two points. This leads to completions, which do
not appear in the first-order case, and do sometimes require more care. But issues of
this type will not concern us here.
S. And what is the gain in moving to continuous logic?
L. As you know, unlike the ‘total’ foundations of mathematics envisaged by
Russell and Whitehead, or Hilbert, model theory has to tame one area at a time;
using continuous logic permitted a great expansion of the repertoire. See for instance
Ben Yaacov et al. (2008). Once one has continuous logic, Robinson’s phrase, that
model theory is the metamathematics of algebra, can be extended to meaningful
parts of analysis. From another angle, constructions such as the monadic space of
a theory, while first encountered in the discrete setting (Shelah 1990; Lascar 1982;
Kim and Pillay 1997), are more naturally presented in the continuous logic setting.
P. It also makes mathematical logic look much more like the foundations of
physics. In a structure one takes an atomic formula, and upon specifying parameters,
obtains a value. In classical physics, we have ‘observables’, often depending on
parameters (such as a test function supported in a small region of some point of
space), and obtains a numerical answer. And a given experiment will confirm a
theory, or measure the value of an observable or a fundamental constant, only to a
given degree of approximation.
L. For many years I was puzzled by the fact that the simplest physical models,
and with them many areas of mathematics, seem not to be approachable by first-
order logic, without the intermediary of a large measure of set theory. The latter
seems to be used only as a scaffolding, yet it makes model-theoretic decidability
results impossible. It is surprising to what extent this impression was due to no
more than a slight discrepancy in the formalization, correctible without any change
of the superstructure it supports.
P. I wonder to what extent the role of Hilbert in the formalization of both
mathematical logic and mathematical physics has to do with this parallelism, later
broken and, I see, mended.8

8 Corry (2004) has an interesting account of the proximity of the two in Hilbert’s own thought.
Hilbert’s foundations of specific subjects looks on this account very much like a direct predecessor
of the model-theoretic program of finding ‘tame’ axiomatizations of wide (but circumscribed) areas
of mathematics; while his ‘continuity axiom’ in this context looks like a part of a description of
continuous logic: If a sufficiently small degree of accuracy is prescribed in advance as our condition
for the fulfilment of a certain statement, then an adequate domain may be determined, within which
one can freely choose the arguments [of the function defining the statement], without however
deviating from the statement, more than allowed by the prescribed degree. (Hilbert 1905, cited in
Corry (2004).)
17 On the Descriptive Power of Probability Logic 383

S. But still, in the specific setting of probability logic, can I just ignore this issue
if I pretend that probabilities are meaningful only to a given precision (say to a
single atom in a mole)? Can I then use a finite coding of probabilities by ordinary
first order formulas?
L. Yes, if you make this assumption you can ignore today’s discussion. When
probabilities behave discretely, they can be treated using ordinary predicates.
Conversely, ordinary quantifiers are clearly special cases of probability quantifiers,
assuming there is a gap between 0 and the next highest possible probability for a
given question. Yet the results we will discuss are already meaningful, even under
such restrictions.9
P. But even if we ignore real-valued logic, taking all probabilities to be rational,
this discussion was useful in order to remind us of the significance of the notion of
approximation, which is surely just as important in discrete first order logic.
I think it underlies most achievements of modern logic, in Gödel’s proof of the
consistency of the continuum hypothesis no less than the Ax-Kochen asymptotic
theory of p-adic fields, to choose two random examples.
L. I agree. Is this notion as pervasive in the philosophy of physics?
P. I think physicists never evoke a mathematical structure without being aware
of the meaning of approximations to the same. They may exhibit this awareness in
their practice, rather than in their definitions, which is why Dirac felt comfortable
with his δ-functions long before Schwarz.
It is also the standard model of the history of physics, as opposed to physics
proper. We have a succession of theories, approximating reality with respect to more
and more classes of phenomena, and to better and better accuracy. Until perhaps one
reaches a final theory, and needs only to fine-tune some parameters, approaching
their physical value.
L. Sometimes it seems to go the other way: the perfect mathematical model is
the theory; the universe is the rough approximation; it is understandable precisely
because it approximates a beautiful theory. A hill of sand has the perfect shape
predicted by a differential equation, whose terms are dictated by considerations
of weight and forces of the wind. But we know that the grains are discrete; the
differential equation (taken literally) is wrong. The right mathematical statement
is that the actual beach lies on a sequence of ideal ones, with sand made of ever
smaller grains; it is the limit shape of the various sand hills, that will really satisfy
the differential equation; reality is just a very good approximation to the ideal.
S. What notion of ‘limit shape’ do you have in mind here?
P. Perhaps some variation on Gromov-Hausdorff distance. In a known ambient
space, say 3-dimensional Euclidean space, Hausdorff showed how to view the space

9 There are in fact interesting mathematical worlds where probabilities are always rational, and
with bounded denominator in any given family. Thus for instance Chatzidakis et al. (1992) study
‘nonstandard finite fields’, infinite models of the theory of all finite fields; they carry a notion of
probability deriving from the counting-based measure on finite fields. There, the probability of
being a square is exactly 12 , as is easy to see; the probability of x and x + 1 both being squares is a
little more complicated to analyze, but still equals exactly 14 .
384 E. Hrushovski

of all compact subsets as a metric space. Gromov defined a metric on abstract

compact metric spaces, given without any embedding in an ambient space. For
instance, any compact metric space X is the limit of finite metric spaces; one can
pick a finite -dense of points, i.e. any point of the space is at distance at most
 from a point of the finite subset; as  approaches zero, the finite subspace will
approach X.
L. Gromov’s distance agrees, incidentally, with the continuous-logic notion
of approximation of structures that we discussed, with respect to existential and
universal formulas.
On the other hand, individual compact metric spaces play a special role in
continuous logic; they are the analogues of finite sets in the discrete case. To any
given resolution, they do look finite.

17.3 Probability Logic

Probabilities behave formally like a quantifier. Given a formula φ, with free variable
x and possibly some others y, one can consider another statement, asserting that the
probability – if one chooses x at random – that φ(x, y) be true, is α. This can be
viewed as a true/false statement about y and α, but is more conveniently construed
(as discussed above) as a statement about y, whose value is α; we can denote it as
Ex φ(x, y).
P. But how do you iterate quantifiers in this language? After all, the complexity
of first-order logic is created by quantifier alternation. And here you started with a
0/1-valued formula, and obtained a real-valued one.
L. It would be possible, though clumsy, to stick with 0/1-valued statements, and
then iterate them, in expressions such as: if we choose x at random, with probability
>.99, a further random choice of y will have probability <.01 of making φ true.
But it is more convenient to generalize probabilities to expectations. For a 0/1-
valued statement, the expectation of the truth value is the same as the probability.
But expectations make sense for a real-valued formula as well, and one can simply
write Ey (Ex φ(x, y)). These two formalisms convey the same information, one can
translate one into the other.
P. I see; we have a formal operator taking a formula φ(x, y, z) (in free variables
x, y, z say) to a formula Ex φ(x, y, z) with free variables y, z – indeed a kind
of quantifier. We view φ(x, b, c) as describing an event relative to the variable
x; depending on parameters b, c. Then Ex φ(x, b, c) gives the probability. Or
more generally, if φ(x, b, c) denotes a ‘weighted event’, or random variable, then
Ex φ(x, b, c) should be the expectation. The first axioms for “fields of probability”
in Kolmogorov (1950) are entirely algebraic; they already occur in some form in
Hilbert and earlier, cf. Corry (2004). It should be easy to formalize them here.
17 On the Descriptive Power of Probability Logic 385

L. Just so. Generalized to expectation of ‘weighted’ events, the axioms read:

• Ex (φ + ψ) = Ex (φ) + Ex (ψ)
• Ex (ψ(y) · φ(x, y)) = ψ(y)Ex (φ(x, y))
• Ex (|φ|) ≥ 0
• Ex (1) = 1
The first three axioms express linearity and positivity of expectation. The fourth is a
normalization axiom, asserting that the probability of a tautologically true condition
is 1 (here the first 1 denotes the constant function with value 1, in various sets of
variables.)
P. But what about countable additivity, the axiom Kolmogorov adds in the second
chapter? All of measure theory appears to depend on this.
L. In our setting, countable additivity comes of itself. Unlike the Kolmogorov
setting where the field of events is an arbitrary Boolean algebra of sets, here
the events are given by formulas. Using a syntax/semantics duality – Stone’s or
Gelfand’s, according to whether the logic is discrete or ‘continuous’ – we can
represent formulas as continuous functions on a compact Hausdorff space (the space
of types). The operator Ex then induces a unique countably additive probability
measure on the algebra of Borel subsets. (Riesz representation theorem.)
P. Great; so this amounts to a completeness theorem for the logic. A consistent
theory has a model, which is a structure in the usual sense along with a probability
measure on the σ -algebra generated by the definable sets.
S. But does it mean that Fubini’s theorem comes for free, too? If we also have
expectation operators for another variable y, is it automatic that Ex Ey φ(x, y) =
Ey Ex φ(x, y)?
L. This is interesting; it speaks again to the difference between unary and
relational. The usual formulation of Fubini’s theorem concerns the product measure;
it measures only sets (or for us formulas) of the form ψ(x) or θ (y), and their
combinations. But φ(x, y) may not arise from unary formulas at all. Thus while
Fubini’s theorem – being a theorem – does hold, it does not give the commutativity
that you mentioned. If you are interested, see Simon (2015), Theorem 7.29 for
example, for a place where such commutativity, or non-commutativity, has model-
theoretic significance.
I realized just today the close connection of what I’m calling here pure
probability logic, with the flag algebras of Razborov (2007). They include additional
axioms – most likely Fubini will suffice- as they are aimed at proving statements of
additive combinatorics. As our interest is in a different question – finding a ceiling
on new concepts definable in this – we are best with the minimal set of axioms. I
strongly recommend by the way the beautiful exposition (Razborov 2013).
P. It is interesting to note non-commutativity of operators, the source of the
uncertainty principle in some presentations, can arise from classical logic for purely
model theoretic reasons.
L. We agreed not to discuss quanta today; but since you raise the analogy, let me
also indulge in a comment on the non-measurability hypothesis of Pitowsky (1982).
I used to think that non-measurable sets are pathological and unlikely to occur in
386 E. Hrushovski

physics; the problem being not doubts about their existence, dependent on the axiom
of choice, but the fact that this axiom, when used to show existence of one set, tends
to bring with it a plethora of others, too similar to each other; such sets can rarely if
ever be uniquely characterized, so that hopes of a meaningful physical theory using
them seem forlorn in advance. But thinking about measures on Cartesian products
suggests a quite different perspective. There, non-measurability only means, in usual
usage: with respect to the product measure. It can simply indicate a binary relation
that does not reduce to unary ones.
Consider a family of finite structures, Fn approaching a structure F∗ . Each
Fn carries a probability measure induced by counting – a subset X has measure
|X|/|Fn |. As probability logic structures, they approach (F∗ , μ1 ) for some measure
μ1 . Likewise, Subsets R of the Cartesian plane F∗2 are always measurable for an
appropriate limit measure μ2 on F∗2 . But they are very rarely measurable for the
product measure μ1 × μ1 .
S. Are these exotic sets that are not measurable? Or everyday ones like the
diagonal?
L. The diagonal on F∗ itself, the graph  of the equality relation, will be
measurable since it has measure zero. But thickenings of the diagonal – pairs of
elements at distance ≤  from each other, for some  and an appropriate notion of
distance – will usually fail to be measurable. In case Fn is a group, there are natural
neighborhoods U of  of the form xy ∈ v ∈ U , where U is a ‘neighborhood’ of 1,
of small but positive measure on Fn . If Fn is a sufficiently noncommutative group,10
such neighborhoods will not be measurable. But there is nothing pathological about
them.
The specific case of non-measurable spin functions on a sphere may not be
attainable in a pseudo-finite manner, as suggested by results in Farah and Magidor
(2012). But the larger idea of a significant two-dimensional set, not measurable
for the product measure but restricting to a measurable set in every linear one-
dimensional section, attains in this light a different allure.
S. But let us go back to completeness. Is there also an incompleteness theorem?
Of course if one throws in the usual quantifiers, the usual incompleteness theorems
apply. But if we use exclusively probabilistic quantifiers, replacing “all” by“almost
all” in the appropriate sense? Gödel showed that an alternation of n quantifiers does
not reduce to an alternation of n − 1; is it likewise true here that alternating “almost
surely” and “almost never”?
P. I would guess not. It is often pointed out (by Laplace as well as Keynes, if
memory serves me) that the knowledge of highly probable events – with probability
very close to 1 – suffices to determine all probabilities; this, by the law of large
numbers, provided we can discuss the probability outcome of a large number
of choices. The same principle can be used to read out, by measuring a single
probability, when (for instance) 99% of elements x there are about 3% of elements

10 I.e.
without a commutative subgroup of bounded index. The more precise statement is that the
map (x, y) → (x, yx) will not be measurable at the limit.
17 On the Descriptive Power of Probability Logic 387

y satisfying some relation. L. It suffices to ‘guess’ a very large number of random

elements, and see if this statement holds for frequency among them.
L. Indeed this is Hoover’s normal form; see Keisler (1985). Any sequence of
probability or expectation quantifiers, and connectives, is equivalent to a single
string Ex1 Ex2 · · · Exn upfront, followed by a formula made up using connectives
from atomic ones. It is a kind of universal quantifier-elimination result for proba-
bility logic. Quite amazing to a model theorist, who spends years studying some
special theory of geometric provenance, and trying to eliminate its quantifiers based
on laws specific to that case.
The obverse side of the coin, of course, is that this is a first hint of the limited
expressive power of probability logic.
S. But it does still leave a block of expectation quantifiers at the front, of
arbitrary length. And even with one quantifier, and one binary relation R, couldn’t
we define complicated, intrinsically n-ary relations, just as in the first-order case, by
quantifying over a Boolean combination?
L. Before discussing this, we will need to refine our notion of monadic. But it is
late; shall we continue tomorrow?

17.4 The Recognition of Space

P. You said probability logic is good at discerning spatial structures. Could you give
some examples?
L. Let us start with Gromov’s reconstruction theorem.11 It applies to a language
for probability logic with a distance predicate d(x, y), satisfying the axioms of
a metric space. A model for this is thus a complete metric space M carrying a
measure; we assume moreover that all balls are measurable, with positive measure.
If there are any measure zero balls, we discard them; probability logic will not notice
the difference, since it is blind to measure zero phenomena. Likewise, if M is not
complete, we can just pass to the completion. Then the theorem asserts that any
such model is determined by its theory. It is usually presented in more concrete
terms: if for each n we know the probability distribution for the distances between
n randomly chosen points, then we know the space.
P. And if M carries additional structure?
L. In continuous logic, all predicates are assumed uniformly continuous with
respect to the structural metric. With this assumption, the theorem remains true.
The case of an additional constant is already interesting; it implies that if the same
geometry is visible when you stand at a or at b inside M – for instance, in the ball
of radius 1 around either of them, there is such-and-such a probability for a triangle

11 Gromov (1998), see Vershik (2004).

388 E. Hrushovski

to be obtuse,12 the same probability for a and b; and ditto for all such questions –
then there exists an automorphism of M carrying a to b.
P. So if the geometry looks the same from any vantage point on M, statistically
speaking, then the space is homogeneous; there are automorphisms taking any given
point to any other.
L. Yes. M will be compact then, and a whole body of theorems on the structure
of compact and locally compact groups applies to it; in the present case, the Peter-
Weyl theorem. It means in particular that with respect to any given formula, or finite
number of formulas, M can be quite precisely approximated by finite unions of
compact, finite dimensional, homogeneous Riemannian manifolds. Such manifolds
are very special; with somewhat stronger homogeneity assumptions (on directions),
they can be fully classified. Spheres and tori of various dimensions are good
examples. It is really a vertiginous change of perspective, from the very abstract
situation we started with to these classical geometries.
This phenomenon persists if the measure is infinite, but finite on any ball. In this
case M is locally compact; Euclidean space, say R3 , is the most classical example.
Locally compact groups were the subject of Hilbert’s 5th problem, resolved in
the 1950s, and showing that they are quite close to Lie groups, finite dimensional
groups with a differentiable structure; the simple Lie groups were classified in the
nineteenth century.
S. Going back – can I summarize Gromov’s theorem by saying that M can be
identified internally – that a physicist in M, using probability logic, and able to
sample distances between points, will be able to recognize the geometry M?
P. Prima facie, the theorem asserts only that the universe is determined by the
answers to all stochastic experiments. So your physicist will get closer and closer to
the full description, that does determine M. But there is no guarantee that at some
given time, inside M, a physicist can guess at M as the right answer. At any point,
there may be many other possibilities.
L. That’s right; there need not exist a finite description of M at all, after all. But
suppose M is in fact homogeneous. Our physicists can see that the formulas they
measure have very close values at different points. They can then make the leap
and assume it is homogeneous. Then I expect there will often (say when Aut (M) is
semisimple) be in fact be a unique homogeneous M where a certain finite number
of stochastic questions have the experimentally found answers.
S. But how can the physicists of M guess that this particular sentence –
expressing homogeneity – once gleaned from the statistics, would have such fruitful
consequences? How to distinguish it from others, that presumably also exist, that
lead to nothing special, or else to a completely wild, Gödelian domain?
L. If we knew that, we’d be able to answer the question: What is model theory?
P. Seriously, it is an interesting topic of mathematical philosophy of physics:
which possible geometries are amenable to discovery by their denizens, in the sense

12 Expressed via the sides, i.e. the square of one side is greater than the sum of squares of the other
two.
17 On the Descriptive Power of Probability Logic 389

that a finite number of statements, if seen to have high probability and assumed to
hold absolutely, will characterize the world uniquely? I think physicists are well
aware of such questions (and for related reasons, were disappointed to find large
numbers of Calabi-Yau manifolds.)
L. This leap incidentally, from high probability to always, cannot be justified
within the probability calculus; I don’t know what it means to say that it is likely to
be true, or becomes likelier with more observations; a condition like homogeneity
would seem in any reasonable model to have probability zero, if it has one at
all. Nevertheless this seems to be the kind of leap science is based on. And then,
following the passage from probabilistic to definite rules, a long mathematical
analysis using classical logic, before going back to probabilistic verification.
S. There is also an issue at the opening that we ignored. We assumed that the
distance between two points is directly observable, and with arbitrary precision. But
what if one only has access to everyday distances, not too small or big? Suppose
these physicists have size on the order of 100 units. They can estimate distances
up to the nearest integer unit; by weighing or other means, like Dalton, they can
measure the number of particles in a region a few units wide ball, or at least the
ratio of particle numbers between two such regions. But to describe the geometry
– e.g. to see whether it is Riemannian – requires understanding distances at much
smaller scales. How are they to get this information?
L. This is not unrelated to Hilbert’s fifth problem, mentioned a few moments
ago. The breakthrough on this problem was Gleason’s construction, starting from
a compact neighborhood of the identity in a group, of the tangent space, the
infinitesimal vectors. A modern version with pseudo-finite ramifications is due to
Breuillard et al. (2012). (and in fact this is how I got interested in such questions.)
The description of the metric at small distances, given the statistics of multiplication
at a medium range, is a step in this construction.
But this can be done without any homogeneity assumptions; the most direct
description is due to Lovász and Szegedy (2010), though their setting was slightly
different.13 Suppose, in your scenario, that R(x, y) denotes the relation: x, y are
at most a unit distance from each other. The Lovász-Szegedy metric can then be
written in one line as a probability logic formula:

dLS (x, y) := Ez (|Et (R(t, x)&R(t, z)) − Et (R(t, y)&R(t, z))|)

The first thought would be that x, y are nearby if the unit balls around them are 99%
equal, in terms of the number of particles they contain; there is of course a grain
of truth in this. But what actually works is at one remove: intersect the unit balls
around x and z, and compare this to the result for y and z. If for the vast majority of

13 The theory of graphons includes ideas very close to those presented here. Lovász and his
collaborators begin with measure-theoretic limits of sequences of graphs, then builds a topology.
In the same years, unaware of this, model theorists added measures to their already topological
type and strong type spaces. Many results are represented in the two theories, in different terms;
but a systematic comparison has not been undertaken.
390 E. Hrushovski

particles z, the two numbers are close, then x, y must be close. Even at the pseudo-
finite limit, where the number of particles in a ball is infinite, this metric can be
shown to be compact; and a unit distance ball, in naive units, is still contained in
a finite ball of this metric. This indicates that the new metric must be seeing short-
range distances fairly faithfully; if all distances below 1 were lumped together as in
the original metric, the unit ball would be infinite and discrete, and not compact. The
story goes further than that, but you see that probability quantifiers are a powerful
tool in the recognition of space.

17.5 Unary and Monadic

P. We were able so far to talk without agreeing on the notion of space; just as
mathematicians felt no need for a definition of algorithm, as long as they were
showing many things were algorithmic. But if we are to discuss the capabilities
of probability logic beyond the spatial, we should perhaps agree what we mean by
the latter.
L. The most basic example is a ‘space’ allowing two positions. We could call one
of them Q. the other then is ¬Q. A unary predicate. In, or out.
S. Isn’t it a little artificial to call this a space? Do any mathematical definitions of
‘space’ apply here?
L. Well, you can think of it as a very basic, two-element topological space X0 ,
with points 0 and 1. Given a structure M, the interpretation of Q gives a map M →
X0 . We think of M as spread out over X, with Q lying above 1, and the complement
above 0.
S. But X0 is discrete; you are not really using the topology.
L. For one predicate, that is true. With more unary predicates, we can distinguish
more and more regions in this primitive space, positions that elements may take;
we can subdivide the house into rooms, or delineate additional areas outside it. As
long as the set of possible positions is finite, it is discrete and the topology is not
used. But generally there are infinitely many unary formulas to consider, and then
the best way to keep track of the possible positions is a topology. For instance if we
have infinitely many unary predicates P1 , P2 , · · · with no relations among them –
all positions are possible – then X0 would be the Cantor space of 0/1-sequences,
with 01 · · · coding P1 ∩ (¬P2 ) ∩ · · · . If constraints exist, we will have a certain
subspace of that.
S. And in continuous logic, I suppose, these positions may form a connected
space; the range of values of a unary predicate can be an extended interval.
L. Quite so. In any case it is compact; up to a given resolution, there are finitely
many possibilities.
P. Perhaps this is a good place to disambiguate the word ‘finite’. When a
combinatorialist like Erdős says ‘let n be a number’, n is a finite number; but one
17 On the Descriptive Power of Probability Logic 391

that we view as potentially increasing towards infinity. A theorem on all but finitely
many n would often be satisfactory. In other cases, it is really viewed as fixed.14
L. Well, the space X0 , to any given resolution, is not pseudo-finite; it is not
moving; it’s dead finite. We may well visualize large finite structures, where Q has
n elements and ¬Q some much bigger number m perhaps; at the limit we talk of a
pseudo-finite n. But X0 will still have two.
S. So the position of an element a in X0 is what is called in Chang and Keisler
(1990) the type of a; it determines the value in M of all unary formulas concerning
a. But given two elements a, b, and a relation R, the positions of a and of b in X0
will not necessarily determine whether R(a, b) holds. By ‘spatial’ you just mean
‘unary’, then?
P. The spatial configuration of n points should be determined by the position of
each of them; in this sense, we do view space as monadic, or unary.
L. So far we took ‘monadic’ to be synonymous to ‘unary’. But the literal notion
of unary is not robust. Starting with a purely unary structure and forgetting some
information, we can arrive at one that is not quite unary. Let us go back to our unary
predicate Q; we think of it as specifying one side of some border, some distinction
in the world. You might think a single unary predicate is the least possible structure
a relational universe can have, but it is not: you can easily conceive of a language
able to specify less: it gives the demarcation line, without breaking the symmetry by
giving a name to either side. You simply need a relation E(x, y), true if both x, y
are in the extension of the predicate, or if both are out; an equivalence relation with
two classes. Technically, E(x, y) is a binary relation. But it could have been defined
from the unary predicate Q, if we had it: E(x, y) ⇐⇒ (Q(x) ∧ Q(y)) ∨ (¬Q(x) ∧
¬Q(y)). A world with E as a basic relation carries no more structure than one with
Q, and possibly less: there may be no way to tell apart the two classes.
S. What happens to X0 in this case?
L. Given a model M, we can define X0 as M/E, the set of classes; even without
having a name for one of them. So we still have the map p : M → X0 , giving what
we take for the spatial location of an element. The elements of X0 are called finite
imaginaries; it was precisely for them, and quite in the present setting, that Shelah
introduced his notion of imaginary elements (or T eq ; Shelah 1990).
S. But as a topological space, X0 is just the same as in the case of unary
predicates?
L. Yes, but there is some further information in it, of a different kind: about what
we are not able to name. If E has n classes (n being a limited number), then X0
has n points; some of them may have names, and some not; if we pick one, we may
have a way of naming another. The right ideas to describe the situation were found
by Galois: there is a finite group acting on the set of classes; all classes, or sets of

14 Jordan (1878), introducing his theorem regarding all but finitely many matrix groups of a given
dimension, attempted to introduce the term ‘limited’ (see p. 114), insisting that limité et illimité
. . . ne sont pas synonymes de fini et infini”, but unlike the concept and theorem, this linguistic
advice was not fully taken up by later mathematicians.
392 E. Hrushovski

classes, or even relations on the classes describable in the language are preserved by
the group, and conversely any invariant relation can be described by a formula. It is
exactly the situation with the roots of an irreducible polynomial; the Galois theory
underlying Shelah’s finite imaginaries directly generalizes the Galois’s theory for
solutions of algebraic equations.
P. And if we use continuous logic?
L. The need for a more general monadic space actually arose earlier, in the
ordinary, two-valued first-order case, in Lascar (1982) and Kim and Pillay (1997).
In place of maps p : M → X0 with X0 a finite set, we have to take into account
maps into compact spaces; but using the Peter-Weyl theorem, it suffices in fact
to consider maps p : M → Y into finite dimensional Riemannian manifolds.
The Galois theory persists: there is a canonical group H acting on Y , a subgroup
of the orthogonal group of automorphisms of Rn preserving Euclidean distance.
A Riemannian distance ρ(y, y  ) on Y is preserved by H . And while the literal
spatial position p(m) in Y is not definable, the distance between two points –
ρ(p(m), p(m ))- is definable in the sense of continuous logic. The full monadic
space X is X is the projective limit of all such p; inasfar as any finite number
of formulas is concerned, to a given approximation, it suffices to think of one
p : M → Y as above.
P. The move you describe from unary to monadic closely parallels the relativistic
correction of Galileo. The elements of the universe are mapped into a space Y ,
but only relative distances are meaningful. And for any given question, if you fix
a sufficient number of reference points, spatial location becomes unary again; it is
determined by the distances to each of these reference points.
S. But the Galilean group is not compact; it includes the translation group R3 ;
like the space R3 itself, it is only locally compact.
L. We mentioned earlier are ‘unbounded’ generalizations of continuous logic;
there is also a natural generalization of probability logic, where the measures are
finite only locally. This should really be our framework, but as they do add a
technical layer, it seemed better to describe the bounded case. In the full setting
we do obtain locally compact groups and spaces; so R3 is included.
P. It is quite striking how Euclidean spaces, that one thinks of as essential but
possibly contingent features of physics and geometry, arise of themselves in a first-
order logic setting. Do you also see something resembling time?
L. Not so far. Nor Lorentzian metrics.
P. I think we can provisionally accept this as the monadic, and in a sense the
compact or locally compact spatial part of the theory of M;
Now what about probability logic?
L. If we have probability quantifiers too, we obtain an induced measure on
the space X; and we can ignore any open set of measure zero, as the probability
quantifiers will not see it. This is quite close to the measured metric spaces we
encountered yesterday.
17 On the Descriptive Power of Probability Logic 393

17.6 Nothing But Space

S. Is it time now to repeat my question from day 3? Suppose we have at least one
binary relation R in the language, two-valued if you like. Define

R ∗ (y1 , . . . , yn ) = (Ex)(R(x, y1 ) ∧ · · · ∧ R(x, yn ))

Isn’t this a new, essentially n-ary relation, obtained in probability logic?

L. Surprisingly, no; in pure probability logic – now including the Fubini axiom
– we can actually write an identity giving the value of R ∗ (b1 , . . . , bn ) in terms
of basic data. If we knew that the events R(x, bi ) are statistically independent,
of course it would just be the product of their probabilities. Now they may not
be independent, since they are spread out differently over the monadic space
X. For instance suppose X has just two points, as above; corresponding to two
classes Q, Q ; say with probabilities 1/3, 2/3, for definiteness. Possibly R(x, b1 )
is contained in Q, and the others are contained in Q ; then of course they are
not independent. A better guess for the expectation of their conjunction would be
(1/3)E((R(x, b1)) + (2/3)E(R(x, b2) ∧ R(x, b3)). The general expression uses
a little more technique from measure theory, but comes to the same thing: we
decompose the sets R(x, bi ) along X, and assume that when conditioned on any
given x ∈ X, they are statistically independent. Then we sum up the result.
S. And does this guess actually work?
L. Yes; provided b1 , . . . , bn do not fork.
S. Do not what?
L. Forking is the fundamental relation found in Shelah (1990), on which the
enormous edifice of stability theory rests. For instance in algebra, an algebraically
independent n-tuple of numbers does not fork; a polynomial relation does cause
forking. What matters to us is that the set of forking tuples has measure zero
(in fact it has measure zero for any definable measure.) In particular the identity
described above holds for almost all b1 , . . . , bn . In pure probability logic, this is
indistinguishable from it holding for all b1 , . . . , bn .
Thus pure probability logic sees, very precisely, the monadic space of the theory;
and conditional probabilities of basic relations as functions on this space. Beyond
this, or relative to it, the universe is boring; everything is locally statistically
independent, no special features can be made out, and no new relations can be
described.
P. This has a Kantian flavor: with probabilistic eyes we see space, since that is
all we are able to see. Classical physics does not take the worlds to be relational;
The dynamics takes place on particles in or fields on this pre-conceived space. It is
sometimes even seen as preliminary to the very language used to formulate queries,
i.e. we may have observables (how many particles in such and such a location?)
whose very definition depends on a point in space. But even if the world were
relational, observed pure probabilistically, it would still appear to be space-based
in a similar way; the binary relation will have no influence on the physical laws, and
will not make itself felt beyond direct, empirical local correlations.
394 E. Hrushovski

S. Do you know Gromov’s (2008), where he speaks about the “reconstruction of

the geometry of physical space” from statistical data?
P. Of course. He first evokes the first genetic map, obtained in Sturtevant (1913)
long before the physical nature of a gene was understood, and confirming at the
same time the one-dimensionality of the genetic code; a beautiful instance of the
statistics-geometry connection. The key is an idea of Morgan’s, that the probability
of a break between A and B should depend on their physical distance. Gromov
proposes a generalization to “geometric structures on a set L that are (. . . ) encoded
by probability measures on the set of all subsets K of L, or by something similar to
such measures.” All definable subsets K in our sense would perhaps be even closer
to the discussion there, of observed samples and arranged experiments. As a model
example, he takes a set L of “pixels (. . . ) or light-sensitive cells in the retina of an
eye (. . . ) stripped of any structure.” L is presented to us as a pure set, with n points;
it comes as part of a two-sorted structure, where the second sort P is the set of all
subsets of L, so of size 2n , and the membership relation is part of the language. An
element of P then codes a black and white picture on L. This is all generic; the
key element is the interpretation of the expectation quantifiers; they reflect not the
counting measure on all of P , but a measure that concentrates on images possible in
a given world, according to their likelihood. Gromov suggests how the geometry
on P may be reconstructed from these statistics: nearby points correlate well,
faraway points not at all. He asks: “ Is there, yet unknown, a mathematical theory
(. . . )incorporating these ideas and being useful not only for “specifying parameters”
in a (given class of) structures but for also for predicting and/or generating new
(classes of) structures?
S. The theory of the monadic space does seem to be a reconstruction of geometry,
incorporating these ideas systematically.
P. Yes; and on it, in this case, the Lovász-Szegedy metric would be natural.
But Gromov may have meant more than this. Measured metric spaces are well-
understood structures; indeed by Gromov’s reconstruction theorem, they can be
parameterized; perhaps we should view the recognition of space, once it is known
that a space is what we are looking for, as merely a form of ‘specifying parameters”?
L. In this case our conclusion is that a mathematical theory such as Gromov
wishes, cannot be found within pure probability logic.
P. But perhaps at the limits of this logic. The forking relation we encountered
earlier is invisible to the probability logic, but is found precisely by tracing the
limits of validity of this logic. Finite improbabilities were incorporated in the
monadic space X, and influenced its form; notably they were used to recognize new
imaginaries, forming another layer of X. Now relative to the full X, none remain. If
any surprises are left – failures of statistical independence relative to X – they are an
infinitely improbable thing; having measure zero they may hint at a new structure.15
Presumably, the next step, for the physicists of L, would be trying to test the new
relation for universal laws, in first order logic again.

15 The discontinuity points of the correlation functions, also of measure zero, may be another such
hint.
17 On the Descriptive Power of Probability Logic 395

L. Let us take a simple example, consisting of addition alone (whereas the

relational language associated with Euclidean geometry, according to Descartes,
has multiplication too.) The set of pixels L, of illimited size n = |L|, is presented to
us ‘stripped of all structure’; but it secretly carries the structure of an abelian group.
To simplify further, assume (L, +) has exponent 2, so it can be viewed as a vector
space over the two-element field F2 = {0, 1}. Let P , the set of pictures, be the dual
group of characters of L; in our case it is the dual space L∗ of homomorphisms from
L to the 2-element field. Each character a ∈ P is thus a black-and-white picture on
L (with a(l) = 1 denoting that l is black.) We treat P using probability quantifiers,
with the unbiased measure on P . As for L, we have a choice, either using probability
quantifiers (with the counting measure) or ordinary ones. In the former case, we will
be unable to discern any structure whatever on L. If we choose any a1 , a2 , a3 ∈ L,
for instance, then with probability 1, all 8 possible pictures on the 3-element set
{a1 , a2 , a3 } will be equally likely, with expectation 2−3 each; nothing to report. But
if we use first-order logic on L, we notice some exceptions. For a small number of
triples – a measure zero set – the probabilities are radically different: the probability
of x(a1 ) = x(a2 ) = x(a3 ) = 0 is 1/4, while x(a1 ) = x(a2 ) = x(a3 ) = 1 has
probability 0. An external observer aware of the addition that secretly rules this
world will see that this happens just when a1 + a2 = a3 . Addition involves forking,
and is inaccessible to the probability logic. A denizen of L, becoming aware of such
exceptions, will have to begin by noting the non-probabilistic associative law, and
analyze what follows in first-order logic, including perhaps the construction of a
new measure appropriate to the graph of addition.
Perhaps a mathematical investigation is warranted, as to what kind of algebraic
structure can be encountered in this way.
S. And what about higher-order relations? Could it simply be that binary rela-
tions, for pure probability logic, form a limited world, but unlimited expressibility
appears once past this barrier?
L. No. By the methods of Towsner (2011), one can see that all formulas defined
in pure probability logic have the general feature we noticed for monadic ones:
to any given approximation, if you fix a sufficient number of reference points as
parameters, φ becomes unary; the value of φ(a1 , . . . , am ) is determined by certain
unary formulas regarding the ai , using the chosen points as parameters. This does
not reveal the precise geometry of the situation as we have in the binary case; but it
does impose a clear limit.
P. We should call it a day. But before we break off, can you say a little more about
the new issues that are seen when the language includes, say, ternary relations?
L. At least two refinements of the notion of space are needed.
The first is simple. In the ternary case already, the additional structure required
involves an inner space associated to each object, with a compact group acting on it
transitively; only relations invariant under the group action are meaningful, so that
this new space contributes nothing at the level of a single object; but still the full
description, involving relations among several, requires it.
S. This is reminiscent of the basic idea of gauge theories in physics, there too the
first thing beyond space!
396 E. Hrushovski

P. But the details appear different – in physics a bundle is induced on the ambient
space, here it appears as still relational and on the particles. This too might be worth
looking into.
L. The second issue is equality, and higher imaginaries; it is probably cognate to
what the higher topos theory of Lurie (2009) is after. We already had to consider
imaginary elements, introduced by Shelah into model theory in precisely this
setting. They are just ordinary elements (or tuple of elements) a, but with a new
notion of equality, provided by a definable equivalence relation. The imaginary is an
ideal element a/E, representing one or any of the elements of a given equivalence
class C. Now suppose the elements a themselves carry some structure (a), and
all the (a) are isomorphic for a ∈ C. We would like to be able to speak of an
ideal , isomorphic to any concrete embodiment (a). This cannot be done with
ordinary imaginaries, but does form a part of the space that probability quantifiers
can discern. A special case, when G = Aut (N/M), N algebraic over a sort M,
R(g, a, b) if g fixes the fibers about (a, b), was treated in Hrushovski (2012).
It is conceivable that these two considerations are the only new ones needed.

Acknowledgements Warm thanks to Dan Drai for his reading and comments on the text; to Martin
Guttman for a conversation, explaining to me inter alia that monads have no windows; to the
editors, Meir and Orly, for their encouragement and patience.

References

Aristotle. (1906). Prior analytics (A. J. Jenkinson, Trans.) and Posterior analytics (G. R. G.
Mure, Trans.). Oxford: Oxford Translation. Available from: http://classics.mit.edu/Aristotle/
prior.html, http://classics.mit.edu/Aristotle/posterior.html.
Ben Yaacov, I. (2008). Continuous first order logic for unbounded metric structures. Journal of
Mathematical Logic, 8(2), 197–223.
Ben Yaacov, I., & Usvyatsov, A. (2010). Continuous first order logic and local stability. Transac-
tions of the American Mathematical Society, 362, 5213–5259.
Ben Yaacov, I., Berenstein, A., Henson, C. W., & Usvyatsov, A. (2008). Model theory for metric
structures. In Model theory with applications to algebra and analysis (Vol. 2, pp. 315–427).
London mathematical society lecture note series (Vol. 350). Cambridge: Cambridge University
Press.
Breuillard, E., Green, B., & Tao, T. (2012). The structure of approximate groups. Publications
Mathematiques- Institut Des Hautes Etudes Scientifiques, 116, 115–221.
Chang, C.-C., & Keisler, H. J. (1966). Continuous model theory. Annals of mathematics studies
(Vol. 58, pp. xii+166). Princeton: Princeton University Press.
Chang, C. C., & Keisler, H. J. (1990). Model theory (3rd ed.). Studies in logic and the foundations
of mathematics (Vol. 73). Amsterdam: North-Holland Publishing Co.
Chatzidakis, Z., van den Dries, L., & Macintyre, A. (1992). Definable sets over finite fields. Journal
für Die Reine und Angewandte Mathematik, 427, 107–135.
Corry, L. (2004). David Hilbert and the axiomatization of physics (1898–1918). From Grundlagen
der Geometrie to Grundlagen der Physik. Archimedes: New studies in the history and
philosophy of science and technology (Vol. 10, pp. xvii+513). Dordrecht: Kluwer Academic
Publishers.
17 On the Descriptive Power of Probability Logic 397

Farah, I., & Magidor, M. (2012). Independence of the existence of Pitowsky spin models.
arXiv:1212.0110.
Gromov, M. (1998). Metric structures for riemannian and non-riemannian spaces. Boston:
Birkhuaser.
Gromov, M. (2008). Mendelian dynamics and Sturtevant’s paradigm. Geometric and probabilistic
structures in dynamics. Contemporary mathematics (Vol. 469, pp. 227–242). Providence:
American Mathematical Society.
Hrushovski, E. (2012). Groupoids, imaginaries and internal covers. Turkish Journal of Mathemat-
ics, 36(2), 173–198. (Arxiv: math.LO/0603413.)
Jordan, C. (1878). Mémoire sur les équations différentielles linéaires intégrale algébriques. Journal
für Die Reine und Angewandte mathematik, 84, 89–215.
Keisler, H. J. (1985). Probability quantifiers. In J. Barwise & S. Feferman (Eds.), Model-theoretic
logics (pp. 509–556). Perspectives in mathematical logic. New York: Springer.
Kim, B., & Pillay, A. (1997). Simple theories. Joint AILA-KGS model theory meeting (Florence,
1995). Annals of Pure and Applied Logic, 88(2–3), 149–164.
Kolmogorov, A. N. (1950). Grundbegriffe der Wahrscheinlichkeitrechnung. Ergebnisse der Math-
ematik (1933); English translation: New York: Chelsea.
Lascar, D. (1982). On the category of models of a complete theory. Journal of Symbolic Logic,
47(2), 249–266.
Lovász, L., & Szegedy, B. (2010). Regularity partitions and the topology of graphons. An irregular
mind, Szemerédi is 70 (pp. 415–446). Berlin/Heidelberg: János Bolyai Mathematical Society
and Springer-Verlag.
Lurie, J. (2009). Higher topos theory. Annals of mathematics studies (Vol. 170). Princeton/Oxford:
Princeton University Press.
Maxwell, J. C. (1925). Matter and motion, 1877. London: Sheldon Press.
Mumford, D. (2000). The dawning of the age of stochasticity. Mathematics: Frontiers and
perspectives (pp. 197–218). Providence: American Mathematical Society.
Pillay, A. (1996). Geometric stability theory. Oxford logic guides (Vol. 32). New York: The
Clarendon Press/Oxford University Press. x+361. Oxford Science Publications.
Pitowsky, I. (1982). Resolution of the Einstein-Podolsky-Rosen and Bell paradoxes. Physical
Review Letters, 48(19), 1299–1302.
Pitowsky, I. (1999). The limits of computability and the limits of formalization (hebrew) (Vol. 48,
pp. 209–229). Iyyun: The Jerusalem Philosophical Quarterly.
Razborov, A. (2007). Flag algebras. Journal of Symbolic Logic, 72, 1239–1282.
Razborov, A. (2013). What is a flag algebra? Notices of the AMS, 60(10), 1324–1327.
Shelah, S. (1990). Classification theory and the number of nonisomorphic models (2nd ed.). Studies
in logic and the foundations of mathematics (Vol. 92). Amsterdam: North-Holland Publishing
Co.
Simon, P. (2015). A guide to NIP theories. Lecture notes in logic (Vol. 44, pp. vii+156). Cambridge:
Cambridge Scientific Publishers. Association for symbolic logic, Chicago.
Sturtevant, A. H. (1913). The linear arrangement of six sex-linked factors in Drosophila, as shown
by their mode of association. Journal of Experimental Zoology, 14, 43–59.
Towsner, H. (2011). A model theoretic proof of Szemerédi?s Theorem. arXiv:1002:4456v3,
January 2011.
Vershik, A. M. (2004). Random and universal metric spaces. In A. Maass, S. Martínez, & J. San
Martín (Eds.), Dynamics and randomness II. Nonlinear phenomena and complex systems (Vol.
10, pp. 199–228). Dordrecht: Kluwer Academic Publishers.
Zilber, B. I. (1984). Strongly minimal countably categorical theories. II. Sibirskii matematicheskii
zhurnal, 25(3), 71–88.
Chapter 18
The Argument Against Quantum
Computers

Gil Kalai

Dedicated to the memory of Itamar Pitowsky

My purpose in this paper is to examine, in a general way, the
relations between abstract ideas about computation and the
performance of actual physical computers.

— Itamar Pitowsky, “The physical Church thesis and physical

computational complexity” (1990)

Abstract We give a computational complexity argument against the feasibility

of quantum computers. We identify a very low complexity class of probability
distributions described by noisy intermediate-scale quantum computers, and explain
why it will allow neither good-quality quantum error-correction nor a demonstration
of “quantum supremacy.” Some general principles governing the behavior of noisy
quantum systems are derived. Our work supports the “physical Church thesis”
studied by Pitowsky (lyuun Jerus Philos Q 39:81–99, 1990) and follows his vision of
using abstract ideas about computation to study the performance of actual physical
computers.

Keywords Quantum computation · Noise sensitivity and stability · Quantum

error-correction · Physical Church-Turing thesis

18.1 Introduction

My purpose in this paper is to give an argument against the feasibility of quantum

computers. In a nutshell, our argument is based on the following statement:

G. Kalai ()
Einstein Institute of Mathematics, The Hebrew University of Jerusalem, Jerusalem, Israel
Efi Arazy School of Computer Science, Interdisciplinary Center Herzliya, Herzliya, Israel

© Springer Nature Switzerland AG 2020 399

M. Hemmo, O. Shenker (eds.), Quantum, Probability, Logic,
Jerusalem Studies in Philosophy and History of Science,
https://doi.org/10.1007/978-3-030-34316-3_18
400 G. Kalai

Noisy quantum systems will not allow building quantum error-correcting codes needed for
quantum computation.

To support this statement we study the relations between abstract ideas about
computation and the performance of actual physical computers, which are referred
to as noisy intermediate-scale quantum computers or NISQ computers for short.
We need to explain why it is the case that quantum error-correcting codes are
needed to gain the computational advantage of quantum computing, and we also
need to explain why quantum error-correcting codes are out of reach. The crux of
the matter lies in a parameter referred to as the rate of noise. We will explain why
reducing the rate of noise to the level needed for good quantum error-correction
will already enable NISQ computers to demonstrate “quantum supremacy.” We
will also explain why NISQ computers represent a very low level computational
complexity class which cannot support quantum supremacy. This provides a good,
in-principle, argument for the infeasibility of both good quantum error-correcting
codes and quantum supremacy.
We will show where the argument breaks down for classical computation.
Rudimentary classical error-correcting codes are supported by the low-level com-
putational complexity class that describes NISQ computers.
Pitowsky (1990) formulates and studies a strong form of the “physical” Church–
Turing thesis (referred to later as the extended Church–Turing thesis (ECTT))1 ,
namely
whether we can invoke physical laws to reduce a computational problem that is manifestly
or presumably of exponential complexity, and actually complete it in polynomial time.

and considers the then recently discovered model of quantum computers2 as a

possible counterexample to this newly formulated physical Church–Turing thesis.
Very strong support for “quantum supremacy” – the ability of quantum computers
to perform certain tasks that classical computers cannot perform, – came four years
after Pitowsky’s paper, with Peter Shor’s discovery (Shor 1994, 1999) that quantum
computers can factor integers efficiently.
Our argument, which is in agreement with quantum mechanics, supports the
validity of the extended Church–Turing thesis but does not rely on this thesis.
Our argument is based on computational complexity considerations and thus
manifests the vision of Pitowsky of using abstract ideas about computation to draw
conclusions on the performance of actual physical computers, and of integrating
computational complexity into theories of nature. Our computational complexity
considerations refer to very low-level computational classes and rely on definite
mathematical theorems; they do not depend on the unproven conjectures underlying
our view of the computational complexity world, such as the P = NP conjecture.
The novelty as well as a certain weakness of my argument is that it uses asymptotic
insights (with unspecified constants) to draw conclusions on the behavior of small-

1 Pitowsky’s paper attributes the physical Church–Turing thesis to Wolfram (1985).

2 Quantum computers were proposed by Feynman (1982), Deutsch (1985), and others. The idea
can be traced to Feynman’s 1959 visionary lecture, “There is Plenty of Room at the Bottom.”
18 The Argument Against Quantum Computers 401

and intermediate-scale systems and, in particular, on the value of constants that

naively appear to reflect just engineering capabilities.
The argument against quantum computers leads to various predictions on near-
term experiments, and to general principles governing the behavior of noisy
quantum systems. It may shed light on a variety of questions about quantum systems
and computation. We emphasize that the argument predicts the failure of near-
term experimental goals of many groups around the world to demonstrate on NISQ
computers quantum supremacy and good-quality quantum error-correcting codes.
In Sect. 18.2 we will consider basic models of computation and basic insights
about computational complexity. Our argument against quantum computers is
presented and discussed in Sect. 18.3, along with concrete predictions on near-
term experimental efforts. The argument crucially relies on the study by Kalai
and Kindler (2014) of noise stability and sensitivity for systems of non-interacting
bosons. This study is built on the theory of noise stability and noise sensitivity
of Boolean functions developed by Benjamini et al. (1999). Section 18.4 dis-
cusses underlying principles behind the failure of quantum computers and some
consequences. Section 18.5 concludes and Sect. 18.6 is about Itamar. This paper
complements, updates, and presents in nontechnical language some of my earlier
works on the topic of quantum computers; see Kalai (2016a,b, 2018), and earlier
papers cited there.
Note added in proof (Nov. 2019). The very recent paper Arute et al. (2019),
claims experimental demonstration of “quantum supremacy” on a 53-qubit quantum
computer. In contrast, the argument of our paper predicts that this supremacy claim
will not stand.

18.2 Basic Models of Computation

Pitowsky (1990) described three types of computers where each type is more
restrictive than the previous one: “the Platonist computer,” “the constructivist
computer,” and “the finitist computer.” We will consider versions of the last two
types. Our definitions rely on the notion of Boolean circuits.3

18.2.1 Pitowsky’s “Constructivist Computer,” Boolean

Functions, and Boolean Circuits

The ‘constructivist computer’ is just a (deterministic) Turing machine, and associated with
this notion is the class of all Turing machine computable, i.e., recursive functions. Can a
physical machine compute a non-recursive function? Itamar Pitowsky (1990)

3 The precise relation between the Turing machine model (considered in Pitowsky 1990) and the
circuit model is interesting and subtle but we will not discuss it in this paper.
402 G. Kalai

Model 1 – Computers (circuits) A computer (or circuit) has n bits, and it can
perform certain logical operations on them. The NOT gate, acting on a single bit, and
the AND gate, acting on two bits, suffice for the full power of classical computing.
The outcome of the computation is the value of certain k output bits. When k = 1
the circuit computes a Boolean function, namely, a function f (x1 , x2 , . . . , xn ) of n
Boolean variables (xi = 1 or xi = 0) so that the value of f is also 0 or 1.

18.2.2 Easy and Hard Problems and Pitowsky’s “Finitist

Computer”

The ‘finitist computer,’ is a deterministic, polynomial-time Turing machine (with a fixed

finite number of tapes). Associated with this notion is the class of all functions which are
computable in a number of steps bounded by a polynomial in the size of the input. As is well
known, many important computational problems seem to lie outside the class P, typically
the so-called NP-hard problem. Itamar Pitowsky (1990)

While general circuits can compute arbitrary Boolean functions, matters change
when we require that the size of the circuit be at most a polynomial in the number
of input variables. Here the size of the circuits is the total number of gates.
Model 2 – Polynomial size circuit A circuit with n input bits of size at most
Anc . Here A and c are positive constants.
The complexity class P refers to problems that can be solved using a polynomial
number of steps in the size of the input. The complexity class NP refers to problems
whose solution can be verified in a polynomial number of steps. Our understanding
of the computational complexity world depends on a whole array of conjectures: NP
= P is the most famous one. Many natural computational tasks are NP-complete.
(A task in NP is NP-complete if a computer equipped with a subroutine for solving
this task can solve, in a polynomial number of steps, every problem in NP.)
Let me mention two fundamental examples of computational “easiness” and
“hardness.” The first example is: multiplication is easy, factoring is hard!4 Multi-
plying two n-digit numbers requires a simple algorithm with a number of steps that
is quadratic in n. On the other hand, the best-known algorithm for factoring an n-
digit number to its prime factors is (we think) exponential in n1/3 (Fig. 18.1).
The second fundamental example of complexity theory is that “determinants are
easy but permanents are hard.” Recall that the permanent of an n-by-n matrix M
has a similar (but simpler) formula compared to the more famous determinant. The
difference is that the signs are eliminated, and this difference accounts for a huge
computational difference. Gaussian elimination gives a polynomial-time algorithm

4 Factoringbelongs to NP ∩ coNP and therefore, according to the basic conjectural view of

computational complexity, it is not an NP-complete problem. The hardness of factoring is a
separate important computational complexity conjecture.
18 The Argument Against Quantum Computers 403

Fig. 18.1 The (conjectured)

view of some main
computational complexity
classes. The red ellipse
represents efficient quantum
algorithms. Note the position
of three important
computational tasks:
factoring, computing
permanents, and boson
sampling

to compute determinants. By contrast, computing permanents is harder than solving

NP-complete problems.5
Classical circuits equipped with random bits lead to randomized algorithms,
which are both practically useful and theoretically important.
Model 3 – Randomized circuit A circuit with n bits of input and with one
additional type of gate that provides a random bit. Again we assume that the circuit
is of size at most Anc (here, again, A and c are positive constants). This time the
output is a sample from a probability distribution on strings of length k of zeroes
and ones.6

18.2.3 Quantum Computers

We will now give a brief description (based on circuits) of quantum computers.

Model 4 – Quantum computers
• A qubit is a piece of quantum memory. A qubit is a unit vector in C2 , and the state
of a qubit is a unit vector in a two-dimensional complex Hilbert space H = C2 .
The memory of a quantum computer (quantum circuit) consists of n qubits and
the state of the computer is a unit vector in the 2n -dimensional Hilbert space, i.e.,
(C 2 )⊗n .

5 The hardness of computing permanents defines an important computational class #P (in words,
number-P or sharp-P) that is larger than PH (the entire “polynomial hierarchy”).
6 For simplicity, we use in this paper the term P (rather than BPP) to describe the computational

power of polynomial-time computers (and polynomial-size circuits) with randomization. We also

use the term Q rather than BQP (for decision problems) and Quantum Sampling (for sampling
problems) to describe the computational power of polynomial-time quantum computers (and
polynomial-size quantum circuits).
404 G. Kalai

• A Quantum gate is a unitary transformation. We can put one or two qubits through
gates representing unitary transformations acting on the corresponding two- or
four-dimensional Hilbert spaces. As for classical computers, there is a small list
of gates that are sufficient for the full power of quantum computing.
• Measurement: Measuring the state of k qubits leads to a probability distribution
on 0–1 vectors of length k.
Here and later we assume that the number of gates is at most polynomial in
n. Quantum computers allow sampling from probability distributions well beyond
the capabilities of classical computers (with random bits). Shor’s famous algorithm
shows that quantum computers can factor n-digit integers efficiently, in roughly n2
steps! It is conjectured that quantum computers cannot solve NP-complete problems
and this implies that they cannot compute permanents.
For the notions of noisy quantum computers and for fault-tolerant quantum
computation that we discuss next, it is necessary to allow measurement throughout
the computation process and to be able to feed the measurement results back into
later gates.

18.2.4 Noisy Quantum Circuits

Next we consider noise. Quantum systems are inherently noisy; we cannot accu-
rately control them, and we cannot accurately describe them. In fact, every
interaction of a quantum system with the outside world amounts to noise.
Model 5 – Noisy quantum computers A noisy quantum circuit has the property
that every qubit is corrupted in every “computer cycle” with a small probability t,
and every gate is t-imperfect. Here, t is a small constant called the rate of noise.
Here, in a “computer cycle” we allow several non-overlapping gates to perform
in parallel. We do not specify the precise technical meaning of “corrupt” and “t-
imperfect,” but the following intuitive explanation (for a restricted form of noise
called depolarizing noise) could be useful. When a qubit is corrupted then its state
is replaced by a uniformly distributed random state on the same Hilbert state. A gate
is t-imperfect if with probability t the state of the qubits that are involved in the gate
is replaced by a uniformly random state in the associated Hilbert space.
Theorem 1 (“The threshold theorem”) If the error rate is small enough, noisy
quantum circuits allow the full power of quantum computing.
The threshold theorem was proved around 1995 by Aharonov and Ben-Or (1997),
Kitaev (1997), and Knill et al. (1998). The proof relies on quantum error-correcting
codes first introduced by Shor (1995) and Steane (1996).
A common interpretation of the threshold theorem is that it shows that large-scale
quantum computers are possible in principle. A more careful interpretation is that
if we can control noisy intermediate-scale quantum systems well enough, then we
can build large-scale universal quantum computers. As we will see, there are good
reasons for why we cannot control the quality of noisy intermediate-scale quantum
systems well enough.
18 The Argument Against Quantum Computers 405

18.2.5 Quantum Supremacy and NISQ Devices

Model 6 – Noisy intermediate-scale quantum (NISQ) computers These are

simply noisy quantum circuits with at most 500 qubits.
The number 500 sounds arbitrary but there is a reason to regard NISQ circuits
and other NISQ devices as an important separate computation model: NISQ circuits
are too small to apply quantum error-correction for quantum fault-tolerance, since
a single good-quality “logical” qubit requires a quantum error-correcting code
described by a NISQ circuit with more than a hundred qubits.
A crucial theoretical and experimental challenge is to understand NISQ comput-
ers.7 Major near-future experimental efforts are aimed at demonstrating quantum
supremacy using NISQ circuits and other NISQ devices, and are also aimed at using
NISQ circuits to build high-quality quantum error-correcting codes.
For concreteness we mention three major near-term experimental goals.
Goal 1: Demonstrate quantum supremacy via systems of non-interacting bosons.
(See Section 18.3.3.)
Goal 2: Demonstrate quantum supremacy on random circuits (namely, circuits
that are based on randomly choosing the computation process, in advance), with
50–100 qubits.
Goal 3: Create distance-5 surface codes on NISQ circuits that require a little over
100 qubits.
In our discussion we will refer to the task of demonstrating with good accuracy
boson sampling for 10–20 bosons as baby goal 1, to the task of building with good
accuracy random circuits with 10–30 qubits, as baby goal 2, and to the task of
creating distance-3 surface codes on NISQ circuits that require a little over 20 qubits,
as baby goal 3.

18.3 The Argument Against Quantum Computers

18.3.1 The Argument

In brief, the argument against quantum computers is based on a computational com-

plexity argument for why NISQ computers cannot demonstrate quantum supremacy
and cannot create the good-quality quantum error-correcting codes needed for
quantum computation. We turn now to a detailed description of the argument, which
is based on the following assertions on NISQ devices.

7 JohnPreskill coined the terms “quantum supremacy” (in 2012) and “NISQ systems” (in 2018).
Implicitly, these notions already play a role in Preskill (1998). The importance of multi-scale
analysis of the quantum computer puzzle is emphasized in Kalai (2016a).
406 G. Kalai

Fig. 18.2 NISQ circuits are computationally very weak and therefore unlikely to create quantum
error-correcting codes needed for quantum computers

(A) Probability distributions described (robustly) by NISQ devices can be

described by low-degree polynomials (LDP). LDP-distributions represent a
very low-level computational complexity class well inside bounded-depth
(classical) computation.
(B) Asymptotically low-level computational devices cannot lead to superior com-
putation.
(C) Achieving quantum supremacy is easier than achieving quantum error-
correction.
Concretely, we argue that Goals 1 and 2 both represent superior computation that
cannot be reached by the low-level computational NISQ devices and that Goal 3
(creating distance-5 surface codes) is even harder than Goal 2 (Fig. 18.2).
We will discuss in Sects. 18.3.3 and 18.3.4 our argument for assertion (A) based
on the theory of noise sensitivity and noise stability, and, in particular, a detailed
study of noisy systems of noninteracting bosons. It would be interesting to explore
other theoretical arguments for (A).
Assertion (B) requires special attention and it can be regarded as both a novel
and a weak link of our argument. There is no dispute that we can apply asymptotic
computational insights to the behavior of computing devices in the small and
intermediate scale when we know or can estimate the constants involved. This
is not the case here. The constants depend on (unknown) engineering abilities to
control the noise. I claim that (even when the constants are unknown) the low-
level asymptotic behavior implies or strongly suggests limitations on the computing
power and hence on the engineering ability. Moreover, these limitations already
apply for the intermediate scale, namely, for NISQ circuits, already when the
number of qubits is rather small. In my view, this type of reasoning is behind many
important successes regarding the interface between the theory of computing and the
18 The Argument Against Quantum Computers 407

practice of computing.8 Several researchers disagree with this part of my analysis

and find it unconvincing.
There is substantial empirical evidence for (C) and experimentalists often refer to
achieving quantum supremacy as a step toward achieving quantum error-correction.
Quantum circuits are expected to exhibit probability distributions beyond the power
of classical computers already for 70–100 qubits, while quantum error-correcting
codes of the quality needed for quantum fault-tolerance will require at least 150–500
qubits. It will be interesting to conduct a further theoretical study of the hypothesis
that the ability to build good-quality quantum error correction codes already leads
to quantum systems that demonstrate quantum supremacy. There is good theoretical
evidence for a weaker statement (C’). (Claims (A), (B), and (C’) already support a
weaker form of the argument against quantum computers.)
(C’) High-quality quantum error-correcting codes are not supported by the very
low-level computational power LDP of NISQ systems.

18.3.2 Predictions on NISQ Computers

The quality of individual qubits and gates is the major factor for the quality of the
quantum circuits built from them. One consequence of our argument is that there is
an upper bound on the quality of qubits and gates, which is quite close, in fact, to
what people can achieve today. This consequence seems counterintuitive to many
researchers. As a matter of fact, the quantum-computing analogue of Moore’s law,
known as “Schoelkopf’s law,” asserts, to the contrary, that roughly every three years,
quantum decoherence can be delayed by a factor of ten. Our argument implies that
Schoelkopf’s law will be broken before we reach the quality needed for quantum
supremacy and quantum fault-tolerance. (Namely, now!)9 Our first prediction is thus
that

8A nice example concerns the computation of the Ramsey number R(k, r), which is the smallest
integer n such that in every party with n participants you can find either a set of k participants where
every two shook hands, or a set of r participants where no two shook hands. This is an example
where asymptotic computational complexity insights into the large-scale behavior (namely, when k
and r tend to infinity) explain the small-scale behavior (namely, when k and r are small integers). It
is easy to verify that R(3, 3) = 6, and it was possible with huge computational efforts (and human
ingenuity) to compute the Ramsey number R(4, 5) = 25 (McKay and Radziszowski 1995). It
is commonly believed that computing R(7, 7) and certainly R(10, 10) is and will remain well
beyond our computational ability. Building gates that are two orders of magnitude better than the
best currently existing gates can be seen as analogous to computing R(10, 10): we have good
computational complexity arguments that both these tasks are beyond reach.
9 Schoelkopf’s law was formulated by Rob Schoelkopf’s colleagues based on Schoelkopf’s

experimental breakthroughs, and only later was adopted as a prediction for the future. An even
bolder related prediction by Neven asserts that quantum computers are gaining computational
power relative to classical ones at a “doubly exponential” rate. Our argument implies that Neven’s
law does not reflect reality.
408 G. Kalai

(a) The quality of qubits and gates cannot be improved beyond a certain threshold
that is quite close to the best currently existing qubits and gates. This applies
also to topological qubits (Sect. 18.3.5).
Our argument leads to several further predictions on near-term experiments, for
example, on distributions of 0–1 strings based on a random quantum circuit (Goal
2), or on circuits aimed at creating good-quality quantum error-correcting codes
such as distance-5 and even distance-3 surface codes (Goal 3).
(b) For a larger amount of noise, robust experimental outcomes are possible but
they will represent probability distributions that can be expressed in terms of
low-degree polynomials (LDP-distributions, for short) that are far away from
the desired noiseless distributions.
(c) For a wide range of smaller amounts of noise, the experimental outcomes
will not be robust. This means that the resulting probability distributions will
strongly depend on fine properties of the noise and hence the outcomes will be
chaotic.
In other words, in this range of noise rate the probability distributions are
still far away from the desired noiseless distributions and, moreover, running
the experiment twice will lead to very different probability distributions.
Following predictions (b) and (c) we further expect that
(d) The effort required to control k qubits to allow good approximations of the
desired distribution will increase exponentially and will fail in a fairly small
number of qubits. (My guess is that the number will be ≤20.)

18.3.3 Non-interacting Bosons

Caenorhabditis elegans (or C. elegans for short) is a species of a free-living round-

worm whose biological study shed much light on more complicated organisms. Our
next model, boson sampling can be seen as the C. elegans of quantum computing.
It is both technically and conceptually simpler than the circuit model and yet it
allows definite and clear-cut insights that extend to more general models and more
complicated situations.
Model 7 – Boson Sampling (Troyansky and Tishby 1996; Aaronson and
Arkhipov 2013) Given a complex n-by-m matrix X with orthonormal rows,
sample subsets of columns (with repetitions) according to the absolute value-
squared of permanents. This task is referred to as boson sampling.
Boson sampling represents a very limited form of quantum computers based on
non-interacting bosons. Quantum circuits can perform boson sampling efficiently
on the nose! There is a good theoretical argument by Aaronson–Arkhipov (2013)
that these tasks are beyond the reach of classical computers. If we consider non-
interacting fermions rather than bosons, we obtain a similar sampling task, called
fermion sampling, with determinants instead of permanents. This sampling task can
be performed by a classical randomized computer (Model 3).
18 The Argument Against Quantum Computers 409

Model 8 – Noisy boson sampling (Kalai and Kindler 2014) Let G be a

complex Gaussian n-by-m noise matrix (normalized so that the expected row norm
is 1). Given an√input matrix A, we average the boson sampling distributions over
√
(1 − t)A + tG. Here, t is the rate of noise.
To study noisy boson sampling we expand the outcome distribution in terms of
Hermite polynomials in nm variables which correspond to the entries of the input
matrix A. The effect of the noise is exponential decay of high-degree terms and, as
it turns out, the Hermite expansion for the boson sampling model is very simple and
quite beautiful (much like the simplicity of the graph of synaptic connectivity of C.
elegans). This analysis leads to the following theorems:
Theorem 2 (Kalai and Kindler 2014) When the noise level is constant, distribu-
tions given by noisy boson sampling are well approximated by their low-degree
Fourier–Hermite expansion. (Consequently, these distributions can be approxi-
mated by bounded-depth polynomial-size circuits.)
Theorem 3 (Kalai and Kindler 2014) When the noise level is higher than 1/n,
noisy boson sampling is very sensitive to noise, with a vanishing correlation between
the noisy distribution and the ideal distribution.
Theorem 2 asserts that noisy intermediate-scale systems of non-interacting
bosons are, for every fixed level of noise, “classical computing machines”; namely,
they can be simulated by classical computers. We do not expect classical computing
devices to exhibit quantum supremacy. In this case, the argument is especially strong
in view of the very low-level computational class described by Theorem 2, as well
as in view of Theorem 3 that asserts that in a wide range of subconstant levels
of noise the experimental outcomes will be unrelated to the noiseless outcomes.
Furthermore, Theorem 3 strongly suggests that in this wide range of subconstant
noise levels, experimental outcomes will be chaotic; namely, the correlation between
the outcomes of two different runs of the experiment will also tend to zero
(Fig. 18.3).

Fig. 18.3 The huge computational gap (left) between boson sampling (purple) and fermion
sampling (green) vanishes in the noisy versions (right)
410 G. Kalai

18.3.4 From Boson Sampling to NISQ Circuits

The following open-ended mathematical conjecture is crucial to part (A) of our

analysis, as well as to predictions (b)–(d) in Sect. 18.3.2.
Conjecture 4 Theorems 2 and 3 both extend to all NISQ computers (in particular,
to noisy quantum circuits) and to all realistic forms of noise.
(i) When the noise level is constant, distributions given by NISQ systems are
well approximated by their low-degree Fourier expansion. In particular, these
distributions are very low-level computationally.
(ii) For a wide range of lower noise levels, NISQ systems are very sensitive to
noise, with a vanishing correlation between the noisy distribution and the ideal
distribution. In this range, the noisy distributions depend on fine parameters of
the noise.
Assertion (A) of our argument is supported by Conjecture 4(i). The conjecture
asserts that, for every fixed level of noise, probability distributions described by
NISQ circuits can be simulated by classical computers. As in the case of boson
sampling, the argument is especially strong in view of the very low-level computa-
tional class described by Conjecture 4(i), and in view of Conjecture 4(ii) that asserts
that in a wide range of subconstant levels of noise the experimental outcomes will
be unrelated to the noiseless outcomes and, furthermore, the correlation between the
outcomes of two different runs of the experiment will also tend to zero.
One delicate conceptual aspect of the argument is that it is a multi-scale
argument. The threshold theorem asserts that noisy quantum circuits for a small
enough level of noise support universal computation! However, in the NISQ
regime the situation is different and Conjecture 4(i) asserts that NISQ circuits are
“classical” and of a very low level of computational complexity even in the classical
computational kingdom. It follows that in the NISQ regime quantum supremacy
(e.g., Goal 2) is out of reach, and high-quality quantum error-correction (e.g. Goal
3) is therefore out of reach as well. This behavior of NISQ circuits implies lower
bounds on the rate of noise that show that, in larger scales, the threshold theorem is
not relevant.

18.3.5 The Scope of the Argument

NISQ computers account for most of the effort to create stable qubits via quantum
error-correction, to demonstrate quantum supremacy, and to ultimately build uni-
versal quantum computers. We expect (Sect. 18.4.1) that noise stability and the very
low-level computational class LDP that we identified for noisy NISQ systems apply
in greater generality.
18 The Argument Against Quantum Computers 411

Specifically, it is plausible that the failure of NISQ computers to achieve

stable logical qubits extends to systems, outside the NISQ regime, that do not
apply quantum error-correction. It is plausible that Google’s attempts to build
“surface codes” with quantum circuits of 50–1000 qubits will meet the same
fate as Microsoft’s attempts to build a single topological qubit based on non-
Abelian anyons. The microscopic quantum process for Microsoft’s attempts is not
strictly in the NISQ regime, but it is plausible that it represents the same primitive
computational power of noisy NISQ circuits.
In brief, the extended argument against quantum computers is based on a
paradox. On the one hand, you cannot achieve quantum supremacy without quantum
error-correction. On the other hand, if you could build quantum error-correcting
codes of the required quality then you could demonstrate quantum supremacy
directly on smaller and simpler quantum systems that involve no quantum error-
correction. To sum
Noisy quantum systems that refrain from using error correction represent a low-level
computational ability that will not allow building quantum error-correcting codes needed
for quantum computation.

It is not easy to define quantum devices that “refrain from using error correction,”
but this property holds for NISQ circuits that simply do not have enough qubits
to use quantum error correction, and appears to hold for all known experimental
processes aimed at building high quality physical qubits including toplogical qubits.
What we see next is that the low level complexity class of NISQ circuits does allow
building basic forms of classical error-correcting codes.

18.3.6 Noise Stability, Noise Sensitivity, and Classical

Computation

The study of noisy boson sampling relies on a study of Benjamini et al. (1999) of
noise stability and noise sensitivity of Boolean functions. See Kalai (2018) where
both these theories and their connections to statistical physics, voting rules, and
computation are discussed. A crucial mathematical aspect of the theory is that
the noise has a simple description in terms of Fourier expansion: it reduces “high
frequency” coefficients exponentially and, therefore, noise stability reflects “low
frequency” (low-degree) Fourier expansion.
The theory of noise stability and noise sensitivity posits a clear distinction
between classical information and quantum information. The class LDP of prob-
ability distributions that can be approximated by low-degree polynomials does not
support quantum supremacy and quantum error-correction but it still supports robust
classical information, and with it also classical communication and computation.
The (“majority”) Boolean function, defined by f (x1 , x2 , . . . , xn ) = 1 if more than
half the variables have a value of 1, allows for very robust bits based on a large
number of noisy bits or qubits and admits excellent low-degree approximations. It
412 G. Kalai

can be argued that every form of robust information, communication, and compu-
tation in nature is based on a rudimentary form of classical error-correction where
information is encoded by repetition (or simple variants of repetition) and decoded
by the majority function (or a simple variant of it). In addition to this rudimentary
form of classical error-correction, we sometimes witness more sophisticated forms
of classical error-correction.

18.3.7 The Extended Church–Turing Thesis Revisited

The extended Church–Turing thesis asserts (informally) that

(ECTT) It is impossible to build computing devices that demonstrate computa-
tional supremacy.
Here, computational supremacy is the ability to perform certain computations far
beyond the power of digital computers.
A classical computing device (or a P-device) is a computing device for which we
can show or have good reason to believe that it can be modeled by polynomial-size
Boolean circuits.
We now formulate the weak extended Church–Turing thesis that reads
(WECTT) It is impossible to build classical computing devices that demonstrate
computational supremacy.
WECTT seems almost tautological and it is in wide agreement. A popular
example of an application of WECCT is to the human brain. As many argue,
it is unlikely that quantum effects in the human brain lead to some sort of
quantum computational supremacy. Another implication of WECCT is that the
recent proposal by Johansson and Larsson (2017) to demonstrate Shor’s factoring
algorithm on certain classical analog devices cannot be realistic.
The argument given in this section asserts that for constant level of noise,
NISQ devices are P-devices. Consequently, WECTT suffices to rule out quantum
supremacy of NISQ devices. What makes this argument stronger is that we have
good evidence that
1. Probability distributions described by NISQ devices for constant error rate are
actually modeled by LDP, a computational class well below bounded-depth
computation, which is well below P.
2. The outcomes of the computation process are chaotic for a wide range of
subconstant levels of errors.
18 The Argument Against Quantum Computers 413

We argue that NISQ devices are, in fact, low-level classical computing devices,
and I regard it as strong evidence that engineers will not be able to reach the level of
noise required for quantum supremacy. Others argue that the existence of low-level
simulation (in the NISQ regime) for every fixed level of noise does not have any
bearing on the engineering question of reducing the level of noise. These sharply
different views will be tested in the next few years.
In his comments on this paper, physicist and friend Aram Harrow wrote that my
predictions are so sweeping that they say that hundreds of different research teams
will all stop making progress. Indeed, my argument and predictions say that the
goals of these groups (such as Goals 1–3 and other, more ambitious, goals) will
not be achieved. Rather, if I am correct, these research teams will make important
progress in our understanding of the inherent limitations of computation in nature,
and the limitations of human ability to control quantum systems.
Aram Harrow also wrote: “The Extended Church–Turing Thesis (ECTT) is a
major conjecture, implicating the work of far more than quantum computing. Long
before Shor’s algorithm it was observed that quantum systems are hard in practice
to simulate because we do not know algorithms that beat the exponential scaling
and work in all cases, or even all practical cases. Simulating the 3 or so quarks
(and gluons) of a proton is at the edge of what current supercomputers can achieve
despite decades of study, and going to larger nuclei (say atomic weight 6 or more)
is still out of reach. If the extended Church–Turing thesis is actually true then there
is a secret poly-time simulation scheme here that the lattice QCD community has
overlooked. Ditto for chemistry where hundreds of millions of dollars are spent
each year simulating molecules. Thus the ECTT is much stronger than just saying
we cannot build quantum computers.”
I agree with the spirit of Harrow’s comment, if not with the specifics, and the
comment can serve as a nice transition to the next section that discusses a few
underlying principles and consequences that follow from the failure of quantum
computing.

18.4 The Failure of Quantum Computers: Underlying

Principles and Consequences

In this section we will propose some general underlying principles for noisy
quantum processes that follow from the failure of quantum computers. (For more,
see Kalai 2016b and Kalai 2018.)
414 G. Kalai

18.4.1 Noise Stability of Low-Entropy States

Principle 1: Probability Distributions Described by Low-Entropy

States Are Noise Stable and Can Be Expressed by Low-Degree
Polynomials

The situation for probability distributions described by NISQ systems appears to be

as follows: noise causes the terms, in a certain Fourier-like expansion, to be reduced
exponentially with the degree. Noise-stable states, namely, those states for which
the effect of the noise is small, have probability distributions expressed by low-
degree terms. Probability distributions realistically (and robustly) obtained by NISQ
devices represent a very low-level computational complexity class, LDP, the class
of probability distributions that can be approximated by low-degree polynomials. In
the absence of quantum fault-tolerance this description extends to (low-entropy)
quantum states that arbitrary quantum circuits and (other quantum devices) can
reach.
Noise stability and descriptions by low-degree polynomials might be relevant
to some general aspects of quantum systems in nature. In the previous section we
considered one such aspect: the feasibility of classical information and computation.
The reason that our argument against quantum computation does not apply for
classical computation is that noise stability still allows basic forms of classical
error-correction. In fact, every form of robust information, communication, and
computation in nature is based on some rudimentary form of classical error-
correction that is supported by low-degree polynomials.
We will now consider two additional aspects: the possibility of efficiently
learning quantum systems, and the ability of physical systems to reach ground states.
Before we proceed I would like to make the following two remarks. The first remark
is that the specific “Fourier-like expansion” required for the notions of noise stability
and noise sensitivity is different for different situations. (The computational class
LDP applies to all these situations.) An interesting direction for future research
would be to extend the noise stability/noise sensitivity dichotomy and to find
relevant Fourier-like expansions and definitions of noise for mathematical objects
of high-energy physics. The second remark is that the study of noise stability is
important in other scientific disciplines, and the assumption that systems are noise
stable (aka “robust”) may provide useful information for studying them. Robustness
is an important ingredient in the study of chemical and biological networks; see,
e.g., Barkai and Leibler (1997).

18.4.1.1 Learnability

It is a simple and important insight of machine learning that the approximate value
of a low-degree polynomial can efficiently be learned from examples. This offers
an explanation for our ability to understand natural processes and the parameters
18 The Argument Against Quantum Computers 415

defining them. Rudimentary physical processes with a higher entropy rate may
represent simple extensions of the class of LDP for which efficient learnability is
still possible, and this is an interesting topic for further research.

18.4.1.2 Reaching Ground States

Reaching ground states is computationally hard (NP-hard) for classical systems,

and even harder for quantum systems. So how does nature reach ground states
so often? The common answer relies on two ingredients: the first is that physical
systems operate in positive temperature rather than zero temperature, and the second
is that nature often reaches meta-stable states rather than ground states. However,
these explanations are incomplete: we have good theoretical reasons to think
that, for general processes in positive temperature, reaching meta-stable states is
computationally intractable as well. First, for general quantum or classical systems,
reaching a meta-stable state can be just as computationally hard as reaching a ground
state. Second, one of the biggest breakthroughs in computational complexity, the
“PCP theorem,” asserts (in physics language) that positive temperature offers no
computational complexity relief for general (classical) systems.
Quantum evolutions and states approximated by low-degree polynomials repre-
sent severe computational restrictions and the results on the hardness of reaching
ground states no longer apply. This may give a theoretical support to the natural
phenomenon of easily reached ground states (Fig. 18.4).

Fig. 18.4 Low-entropy quantum states give probability distributions described by low degree
polynomials, and very low-entropy quantum states give chaotic behavior. Higher entropy enables
classical information
416 G. Kalai

18.4.2 Noise and Time-Dependent Evolutions

Principle 2: Time-Dependent (Local) Quantum Evolutions Are

Inherently Noisy

It is an interesting challenge to give a satisfactory mathematical description of what

“time dependent” means. We proposed in Kalai (2016a) to base a lower bound on
the rate of noise on a measure of the non-commutativity of unitary operations. This
measure of non-commutativity can also serve as an “internal” clock of the noisy
quantum evolution. We also proposed in Kalai (2016a) a restricted class of noisy
evolutions called “smoothed Lindblad evolutions” defined via a certain smoothing
operation that mathematically expresses the idea that in the absence of quantum
error-correction, noise accumulates. An important consequence of Principle 2 is
that there are inherent lower bounds on the quality of qubits, or, more concretely,
an inherent upper bound on the number of non-commuting Pauli operators you can
apply to a qubit before it get destroyed. We note that lower bounds on the rate of
noise in terms of a measure of non-commutativity arise also in Polterovich (2014)
in the study of quantizations in symplectic geometry.

18.4.3 Noise and Correlation

Principle 3: Entangled Qubits Are Subject to Positively

Correlated Noise

Entanglement, the quantum analog of correlations, is a notion of great importance

in quantum physics and in quantum computing, and the famous cat state √1 |00 +
2
√1 |11 represents the simplest example of entanglement and the strongest form
2
of entanglement between two qubits. Principle 3 asserts that the errors for the two
qubits of a cat state necessarily have large positive correlation. This is an observed
and accepted phenomenon for gated qubits and without quantum error-correction
it is inherited to all pairs of entangled qubits. Error synchronization refers to a
substantial probability that a large number of qubits, much beyond the average rate
of noise, are corrupted. An important consequence of Principle 3 is that complicated
states and evolutions lead to error-synchronization. Principle 3 leads to additional
predictions about NISQ systems that can be tested for current experiments.
(e) Every pair of gated qubits will be subject to errors with large positive correla-
tion.
(f) Every pair of qubits in a cat state will be subject to errors with large positive
correlation.
18 The Argument Against Quantum Computers 417

(g) All experimental attempts to achieve Goals 2 and 3 will lead to a strong effect
of error-synchronization.
We emphasize that our argument against quantum computers in Sect. 18.3 does
not depend on Principle 3 and, moreover, the main trouble with quantum computers
is that the noise rate of qubits and gates cannot be reduced toward the threshold for
fault-tolerance. Correlation is part of this picture since, as is well known, the value
of the threshold becomes smaller when there is a larger positive correlation for 2-
qubit gated errors. We note (Kalai 2016a) that Prediction (f) implies that the quality
of logical qubits via quantum error-correction is limited by the quality of gates. See
also Chubb and Flammia (2018) for an interesting analysis of the effect of correlated
errors on surface codes.
A reader may question how it is possible that entangled states will experience
correlated noise, even long after they have been created and even if the constituent
particles are far apart and not directly interacting. We emphasize that Principle 3
does not reflect any form of “non local” noise and that no “additional” correlated
noise on far apart entangled qubits is ever assumed or conjectured. Principle 3
simply reflects two facts. The first fact is that errors for gated qubits are positively
correlated from the time of creation. The second fact is that unless the error rate
is small enough to allow quantum error-correction (and this may be precluded by
the argument of the previous section), cat states created indirectly via a sequence of
computational steps will inherit errors with large positive correlation.
Remark Error-synchronization is related to an interesting possibility (Kalai 2016b)
that may challenge conventional wisdom and can also be checked for near term
NISQ (and other) experiments. A central property of Hamiltonian models of noise
in the study of quantum fault-tolerance (see, e.g., Preskill 2013), is that error
fluctuations are sub-Gaussian. Namely, when there are √ N qubits, the standard
deviation for√the number of qubit errors behaves like N and the probability of
more than t N errors decays as it does for Gaussian distribution. Many other
models in mathematical physics have a similar property. However, it is possible
that fluctuations for the total amount of errors for noisy quantum circuits and other
quantum systems with interactions are actually super-Gaussian and perhaps even
linear.

18.4.4 A Taste of Other Consequences

The failure of quantum computers and the infeasibility of high-quality quantum

error-correction have several others far-reaching consequences imposing severe
restrictions of humans’ and nature’s ability to control quantum states and evolutions.
Following the tradition of using cats for quantum thought experiments, consider an
ordinary living cat.
If quantum computation is not possible, it will be impossible to teleport the cat;
it will be impossible to reverse time in the life of the cat; it will be impossible to
418 G. Kalai

Fig. 18.5 Human imagination allows for a cat with an unusual geometry. Drawing by Netta Kasher
(a tribute to Louis Wain)

implement the cat on a very different geometry; it will be impossible to superpose

the lives of two distinct cats; and, finally, even if we place the cat in an isolated and
monitored environment, the life of this cat cannot be predicted.
These restrictions will already be in place when our “cat” represents realistic
(but involved) quantum evolutions described by quantum circuits in the small scale,
namely, with less than 50 qubits (baby goals 1–3 can serve as good candidates for
such “cats”). See Kalai (2016b) for a detailed discussion (Fig. 18.5).

18.5 Conclusion

The intrinsic study of computation transcends human-made artifacts, and its expanding con-
nections and interactions with all sciences, integrating computational modeling, algorithms,
and complexity into theories of nature and society, marks a new scientific revolution! Avi
Wigderson – Mathematics and Computation 2019.

Understanding noisy quantum systems and potentially even the failure of quan-
tum computers is related to the fascinating mathematical theory of noise stability
and noise sensitivity and its connections to the theory of computing. Our study
18 The Argument Against Quantum Computers 419

shows how insights from the computational complexity of very low-level compu-
tational classes support the extended Church–Turing thesis. Exploring this avenue
may have important implications for various areas of quantum physics. These
connections between mathematics, computation, and physics are characterized by
a rich array of conceptual and technical methods and points of view. In this study,
we witness a delightfully thin gap between fundamental and philosophical issues on
the one hand and practical and engineering aspects on the other.

18.6 Itamar

This paper is devoted to the memory of Itamar Pitowsky. Itamar was great. He
was great in science and great in the humanities. He could think and work like a
mathematician, and like a physicist, and like a philosopher, and like a philosopher
of science, and probably in various additional ways. And he enjoyed the academic
business greatly, and took it seriously, with humor. Itamar had an immense human
wisdom and a modest, level-headed way of expressing it.
Itamar’s scientific path interlaced with mine in many ways. Itamar made
important contributions to the foundations of the theory of computation, probability
theory, and quantum mechanics, all related to some of my own scientific endeavors
and to this paper. Itamar’s approach to the foundation of quantum mechanics was
that quantum mechanics is a theory of non-commutative probability that (like
classical probability) can be seen as a mathematical language for the other laws
of physics. (My work, to a large extent, adopts this point of view. But, frankly, I do
not quite understand the other points of view.) (Fig. 18.6)
In the late 1970s when Itamar and I were both graduate students I remember him
enthusiastically sharing his thoughts on the Erdős–Turán problem with me. This is
a mathematical conjecture that asserts that if we have asequence of integers 0 <
a1 < a2 < · · · < an < · · · such that the infinite series 1
an diverges then we can
find among the elements of the sequence an arithmetic progression of length k, for
every k. Over the years both of us spent some time on this conjecture without much
to show for it. This famous conjecture is still open even for k = 3. Some years later
both Itamar and I became interested, for entirely different reasons, in remarkable
geometric objects called cut polytopes. Cut polytopes are obtained by taking the
convex hull of characteristic vectors of edges in all cuts of a graph. Cut polytopes
arise naturally when you try to understand correlations in probability theory and
Itamar wrote several fundamental papers studying them; see Pitowsky (1991). Cut
polytopes came into play, in an unexpected way, in the work of Jeff Kahn and myself
where we disproved Borsuk’s conjecture (Kahn and Kalai 1993). Over the years,
Itamar and I became interested in Arrow’s impossibility theorem (Arrow 1950),
which Itamar regarded as a major twentieth-century intellectual achievement. He
gave a course centered around this theorem and years later so did I. The study of the
noise stability and noise sensitivity of Boolean functions has close ties to Arrow’s
theorem.
420 G. Kalai

The role of skepticism in science and how (and if) skepticism should be
practiced is a fascinating issue. It is, of course, related to this paper that describes
skepticism over quantum supremacy and quantum fault-tolerance, widely believed
to be possible. Itamar and I had long discussions about skepticism in science, and
about the nature and practice of scientific debates, both in general and in various
specific cases. This common interest was well suited to the Center for the Study of
Rationality at the Hebrew University that we were both members of, where we and
other friends enjoyed many discussions, and, at times, heated debates.
A week before Itamar passed away, Itamar, Oron Shagrir, and I sat at our little
CS cafeteria and talked about probability. Where does probability come from? What
does probability mean? Does it just represent human uncertainty? Is it just an
emerging mathematical concept that is convenient for modeling? Do matters change
when we move from classical to quantum mechanics? When we move to quantum
physics the notion of probability itself changes for sure, but is there a change in the
interpretation of what probability is? A few people passed by and listened, and it
felt like this was a direct continuation of conversations we had while we (Itamar and
I; Oron is much younger) were students in the early 1970s. This was to be our last
meeting.

Acknowledgements Work supported by ERC advanced grants 320924, & 834735, NSF grant
DMS-1300120, and BSF grant 2006066. The author thanks Yuri Gurevich, Aram Harrow, and
Greg Kuperberg for helpful comments, and Netta Kasher for drawing Figs. 18.1, 18.2, 18.3, 18.4,
and 18.5.

Fig. 18.6 Itamar Pitowsky

around 1985
18 The Argument Against Quantum Computers 421

References

Aaronson, S., & Arkhipov, A. (2013). The computational complexity of linear optics. Theory of
Computing, 4, 143–252.
Aharonov, D., & Ben-Or, M. (1997). Fault-tolerant quantum computation with constant error. In
STOC’97 (pp. 176–188). New York: ACM.
Arrow, K. (1950). A difficulty in the theory of social welfare. Journal of Political Economy, 58,
328–346.
Arute, F., et al. (2019). Quantum supremacy using a programmable superconducting processor.
Nature, 574, 505–510.
Barkai, N., & Leibler, S. (1997). Robustness in simple biochemical networks. Nature, 387, 913.
Benjamini, I., Kalai, G., & Schramm, O. (1999). Noise sensitivity of Boolean functions and
applications to percolation. Publications Mathématiques de l’Institut des Hautes Études
Scientifiques, 90, 5–43.
Chubb, C. T., & Flammia, S. T. Statistical mechanical models for quantum codes with correlated
noise, arXiv:1809.10704.
Deutsch, D. (1985). Quantum theory, the Church–Turing principle and the universal quantum
computer. Proceedings of the Royal Society of London A, 400, 96–117.
Feynman, R. P. (1982). Simulating physics with computers. International Journal of Theoretical
Physics, 21, 467–488.
Johansson, N., & Larsson, J.-A. Realization of Shor’s algorithm at room temperature, arX-
ive:1706.03215.
Kahn, J., & Kalai, G. (1993). A counterexample to Borsuk’s conjecture. Bulletin of the American
Mathematical Society, 29, 60–62.
Kalai, G. (2016a). The quantum computer puzzle. Notices of the American Mathematical Society,
63, 508–516.
Kalai, G. (2016b). The quantum computer puzzle (expanded version), arXiv:1605.00992.
Kalai, G. (2018). Three puzzles on mathematics, computation and games. In Proceedings of the
International Congress of Mathematicians, Rio de Janeiro (Vol. I, pp. 551–606).
Kalai, G., & Kindler, G. Gaussian noise sensitivity and BosonSampling, arXiv:1409.3093.
Kitaev, A. Y. (1997). Quantum error correction with imperfect gates. In Quantum communication,
computing, and measurement (pp. 181–188). New York: Plenum Press.
Knill, E., Laflamme, R., & Zurek, W. H. (1998). Resilient quantum computation: Error models and
thresholds. Proceedings of the Royal Society of London A, 454, 365–384.
McKay, B. D., & Radziszowski, S. P. (1995). R(4,5)=25. Journal of Graph Theory, 19, 309–322.
Pitowsky, I. (1990). The physical Church thesis and physical computational complexity. lyuun, A
Jerusalem Philosophical Quarterly, 39, 81–99.
Pitowsky, I. (1991). Correlation polytopes: Their geometry and complexity. Mathematical Pro-
gramming, A50, 395–414.
Polterovich, L. (2014). Symplectic geometry of quantum noise. Communications in Mathematical
Physics, 327, 481–519.
Preskill, J. (1998). Quantum computing: Pro and con. Proceedings of the Royal Society of London
A, 454, 469–486.
Preskill, J. (2013). Sufficient condition on noise correlations for scalable quantum computing.
Quantum Information and Computing, 13, 181–194.
Shor, P. W. (1995). Scheme for reducing decoherence in quantum computer memory. Physical
Review A, 52, 2493–2496.
Shor, P. W. (1999). Polynomial-time algorithms for prime factorization and discrete logarithms
on a quantum computer. SIAM Review, 41, 303–332. (Earlier version, Proceedings of the 35th
Annual Symposium on Foundations of Computer Science, 1994.)
Steane, A. M. (1996). Error-correcting codes in quantum theory. Physical Review Letters, 77, 793–
797.
422 G. Kalai

Troyansky, L., & Tishby, N. (1996). Permanent uncertainty: On the quantum evaluation of the
determinant and the permanent of a matrix. In Proceedings of the 4th Workshop on Physics and
Computation.
Wigderson, A. (2019). Mathematics and computation. Princeton University Press. Princeton, New
Jersey.
Wolfram, S. (1985). Undecidability and intractability in theoretical physics. Physical Review
Letters, 54, 735–738.
Chapter 19
Why a Relativistic Quantum Mechanical
World Must Be Indeterministic

Avi Levy and Meir Hemmo

Abstract We present the framework of Empirical Models and Hidden Variables

Models in the foundations of Quantum Mechanics and its relation to Special
Relativity. In this framework we discuss the definitions of properties such as non-
signaling, strong and weak determinism, locality and, especially contextuality and
show that (under mild assumptions) contextuality is equivalent to non-locality and
to weak determinism, and non-contextuality is equivalent to locality and strong
determinism. The central derivation of this paper is that a contextual Empirical
Model cannot be accounted for by a Lorentz invariant (in the broad sense; see
Sect. 19.1) and deterministic (strong or weak) theory. Hence, a relativistic theory
which yields the empirical predictions of quantum mechanics must be genuinely
stochastic (i.e., cannot be deterministic at its fundamental level). Finally, we discuss
certain debates about the existence of relativistic versions of Bohm’s and the GRW
interpretations of quantum mechanics.

Keywords Quantum contextuality · Quantum probability · Weak-determinism ·

Strong-determinism · Locality · Special relativity

19.1 Introduction

It is well known, as demonstrated by Bell (1964), that quantum probability theory

differs from classical probability theory (see for example Pitowsky 1986, 1989,
2007, for explicit arguments in this direction). In fact, there is a continuum of
probability theories in which the classical non-contextual theory of probability (as
described by Kolmogorov’s 1933 axioms) stands at one end and highly contextual
theories (e.g., the PR box; see Popescu and Rohrlich 1994) stand at the other
end. The contextual quantum probability stands somewhere in between these two

A. Levy · M. Hemmo ()

Department of Philosophy, University of Haifa, Haifa, Israel
e-mail: levya@technion.ac.il; meir@research.haifa.ac.il

© Springer Nature Switzerland AG 2020 423

M. Hemmo, O. Shenker (eds.), Quantum, Probability, Logic,
Jerusalem Studies in Philosophy and History of Science,
https://doi.org/10.1007/978-3-030-34316-3_19
424 A. Levy and M. Hemmo

ends.1 One of the aims of this paper (see more details below) is to give a general
characterization of contextuality in the context of Quantum Mechanics (QM) and its
relation to other concepts such as locality and determinism.
Empirical Models (EM) (defined in Brandenburger and Yanofsky 2008; Abram-
sky and Brandenburger 2011) are probabilistic frameworks for analyzing the
properties of measurement scenarios that arise in both classical and contextual
experiments. One may view the EM framework as a general mathematical frame-
work in which the full scope of probability theories can be analyzed. An EM consists
of a set of admissible measurement sequences and the space of their corresponding
outcomes together with a probability structure defined on these spaces. In simple
terms, an EM is a “table” of measurements and their outcomes with the relative
frequency at which these outcomes are observed.
The more commonly used term Hidden Variables Model (HVM) is an extension
of the EM where a “state” space is appended to the measurement and outcome
spaces. In a HVM the state of the measured system is often taken to provide an
“explanation” for the frequencies of the measurement outcomes captured in the
corresponding EM. The EM is the phenomenological description of the experimen-
tal results while the HVM describes a deeper structure that realizes the observed
measurement outcomes and their relative frequencies as depicted by the EM. Since
in general there may be several HVMs, that differ in their properties, all of which
give rise to the same EM, one may conclude that the EM is the more fundamental
layer of describing the physical reality in the sense that it gives the empirically
compelling constraints on the HVM.
The EM and HVM are very general structures that are not restricted specif-
ically to quantum experiments and can describe any reasonable experimental
setting. As demonstrated beautifully in Brandenburger and Yanofsky (2008) (see
also Abramsky and Brandenburger 2011; Brandenburger and Keisler 2016) the
HVM framework allows defining, in an accurate mathematical way, interesting
properties of physical systems such as Determinism (Strong and Weak), Locality,
No-Signaling and Contextuality and deriving relations between these properties.
(We use these terms now without explicit definitions; we shall give rigorous
mathematical definitions as we proceed.)
Our first aim in this paper is to extend the definitions of the physical properties
of Determinism, Locality and Non-Signaling from the HVM framework to the EM
framework. Such extension is partially proposed in the above-mentioned references,
but we would like to complete it here. If one looks only at the EM, the relations
between these properties can be attributed to the actual “physical reality” rather than
to a specific HVM theory. Indeed, this is exactly the way in which Bell’s Theorem
(1964) is understood (see e.g. Pitowsky 1986; Albert 1992; Maudlin 1994), namely
that any violation of the Bell inequality (given Bell’s assumptions) entails that the

1 The “degree” of contextuality can be measured by the value of certain bounds on inequalities
pertaining to the probability distribution that are violated for contextual (non-classical) probability
theories.
19 Why a Relativistic Quantum Mechanical World Must Be Indeterministic 425

physical reality itself is nonlocal. We shall define the above properties in terms of
EM, examine their relations and compare them to the corresponding properties and
relations of the HVM.
The second aim of this paper is to characterize the contextuality property which
as we will show is relevant to EM only. There seems to be some confusion in the
literature about how precisely to understand contextuality and we would like to
clarify it. This discussion will lead us to propose that under reasonable assumptions
Non-Contextuality is in a sense equivalent to Locality and to Strong Determinism,
and Contextuality is equivalent to Non-Locality and to Weak (that is, not Strong)
Determinism.
Our third aim is to consider the EM and HVM framework in the context of
Special Relativity (SR). The No-Signaling property is an implicit requirement
stemming from the principles of SR, which implies that no physical signal can
be transmitted in a superluminal speed. Here we shall add explicitly the postulate
according to which a HVM compatible with SR must be genuinely Lorentz invariant,
where we understand this last requirement in a broad sense as follows: the laws of
physics should be the same in every Lorentz frame of reference, and moreover there
is no (absolute) temporal order (and therefore no preferred Lorentz reference frame)
that fixes the temporal direction of the cause and effect2 relation. We shall take it
that these conditions should hold in any relativistic theory regardless of whether
or not a violation of them may be empirically detected. We shall call this condition
from now on Lorentz invariance. We will show (in Sect. 19.5) that requiring Lorentz
invariance in this sense implies the No-Signaling property. In fact, assuming Lorentz
invariance, we will prove the following general result:
A contextual Empirical Model does not have a Strongly Deterministic (Local) or
a Weakly Deterministic (Non-Local) Hidden Variable Model.
Since the quantum mechanical probabilities of measurement outcomes that have
been confirmed by actual experiments match a contextual empirical model (i.e.
the quantum probabilities are non-classical; see Sect. 19.4), our proof of the
above proposition implies that under these assumptions (including our additional
requirement of the existence of no preferred Lorentz frame) the physical reality
must be genuinely stochastic. In other words, God must play dice in a relativistic
quantum world.
Bell (1964) was the first to prove that the EPR setting cannot be realized by
a local and (strongly) deterministic HVM. Results pertaining to non-local (weak)
determinism have been proposed in certain other contexts, e.g., Hardy and Squires
(1992), Clifton et al. (1992), Conway and Kochen (2006, 2009), Gisin (2011).
But our theorem above is obtained in a very general context which is explicitly
independent of quantum theory, and which is constrained only by the contextuality
of measurement outcomes.

2 Regardless of how precisely should one understand the relation of causation.

426 A. Levy and M. Hemmo

The paper is organized as follows. In Sect. 19.2 we briefly give the basic
definitions, examples and the main results in the EM and HVM framework. In
Sect. 19.3 we present the definitions of the physical properties mentioned above
and their logical relations in the EM framework. In Sect. 19.4 we analyze the
property of contextuality. In Sect. 19.5 we extend the discussion of EM and HVM
to the framework of Special Relativity and we prove our general theorem, namely
that there is no deterministic theory that is consistent with both the predictions
of quantum mechanics and with the relativistic postulate of no preferred Lorentz
frame of reference. In Sect. 19.6 we briefly draw some implications of our proof
with regard to current interpretations of quantum mechanics. Section 19.7 is the
conclusion.

19.2 The Hidden Variable Model (HVM) Framework

This section is rather technical and uses probability theory for the definitions
part. We put it forward for the sake of completeness of the paper. The attached
examples and the results pertaining to the relations between the properties can
be understood without going into the formal details of all the definitions. Unless
otherwise stated, all the definitions and results, but not the examples, are based
on Brandenburger and Yanofsky (2008), Abramsky and Brandenburger (2011) and
Brandenburger and Keisler (2016). However, our definitions of Empirical Model
and Hidden Variable Model are closer to those in Abramsky and Brandenburger
(2011) than to the definitions in the other two references. Hence there are certain
differences in the main formulations which we shall discuss at the end of Sect.
19.2.2. When appropriate we add explicitly in comments the physical motivation
and interpretation of the formal definitions, as we understand them.

19.2.1 Basic Definitions

Definition: Empirical Model (EM) Let X be a finite set of Measurements

{A, B, C, . . . } and let M be a Measurement Context i.e. a set of subsets of X that
constitutes “admissible series” of measurements.
For example, if X = {A, B, C} then M may be {(A, B), (B, C)} where (A, C) and
(A, B, C) are not included in M.
Remark the “admissible series” of measurement concept originated from QM
where certain series of measurements may be inadmissible as they cannot be
performed simultaneously, since the corresponding operators do not commute.
For each measurement x ∈ X there exists a finite outcome set Ox and hence
for each admissible series m = (m1 , m2 , . . . ) ∈ M, where mi ∈ X, there is an
19 Why a Relativistic Quantum Mechanical World Must Be Indeterministic 427

outcome space Om = Om1 × Om2 × . . . . For example, consider the experiment

of tossing two dice independently from a set of 3 dice, then the set of measurements
may be denoted X = {A, B, C} and for each measurement the outcome set is
Ox = O = {1, 2, 3, 4, 5, 6}.
Let the physical situation be such that it is possible to toss only two dice
simultaneously: A, B or A, C together, but not B, C.
Then the measurement context is M = {(A, B), (A, C)} and Om = {1, . . . , 6}
× {1, . . . , 6} for both members of M.
The Experiments Space E is the set of all admissible measurement series together
with the sets of all their possible outcomes E = {Em = {(m, Om )} : m ∈ M}.
An Empirical Model is the experiments space E together with a family of
probability distributions Qm on the set of possible outcomes of each admissible
series m ∈ M.
In the example of the dice, given that all dice are fair, the experiments space E is
the union of two disjoint subsets:
EAB = {(a, b) : a ∈ {1, . . . , 6}, b ∈ {1, . . . , 6}} and EAC = {(b, c) : a ∈ {1, . . . , 6},
c ∈ {1, . . . , 6}}.
And the probabilities

family)is correspondingly.

) 
QAB = Q (a, b) = 1 36 , QAC = Q (a, c) = 1 36 for a ∈
{1, . . . , 6} , b ∈ {1, . . . , 6} , c ∈ {1, . . . , 6} .
Definition: Hidden Variable Model (HVM) A State Space is a set  whose
elements are denoted by λ.
A hidden variable model is a pair (V, {Pm }) where V = E ×  is the product of
an Experiments Space E with a State Space , and the product is equipped with a
family of probability distributions Pm on the pairs (Em × ).
Definition: A Consistent HVM An HVM (E × , {Pm }) is said to be
consistent with the empirical model (E, {Qm }) if QA, B, . . . (a, b, . . . ) = PA, B, . . .
{(a, b, . . . ) × } for every admissible series of measurements (A, B, . . . ) ∈ M and
every outcome (a, b, . . . ) ∈ OA, B, . . . .
That is, a given HVM is consistent with the EM if one can describe the
probabilities of the measurements outcomes by summing over all the possible
system states λ in the model.
For example, a consistent HVM for the EM presented before of tossing two dice
can be constructed by defining the state space of the three dice as the set of all
three-tuples λ = (α, β, γ ), where α, β, γ ∈ {1, 2, 3, 4, 5, 6}.
The probability distribution that ensures the consistency is given by setting
p (λ) = 613 uniformly and defining:
PA, B (a, b, λ) = PA, B (a, b| λ)p(λ), PA, C (a, c, λ) = PA, C (a, c| λ)p(λ) where
PA, B (a, b| (α, β, γ )) = δ(a = α)δ(b = β), PA, C (a, c| (α, β, γ )) = δ(a=α)δ(c=γ ).
428 A. Levy and M. Hemmo

A,B (a, b) = 36 and direct compu-

1
This is because,
 as we have seen, Q
tation yields PA,B {(a, b) × Λ} = PA,B (a, b| (α, β, γ )) p (α, β, γ ) =
λ∈Λ α,β,γ

δ (a = α) δ (b = β) 613 = 612 .
α,β,γ
)
(and similarly for QAC (a, c) = 1 36 ).

Comments
• Given an EM (E, {Qm }), the HVM defined by (E × , {Pm }), where  is a
singleton state space  = {λ}, and PA, B, . . . (a, b, . . . , λ) = QA, B, . . . (a, b, . . . ) is
consistent with (E, {Qm }). Hence every EM has a “trivial HVM” that is consistent
with it.
• Given an HVM (E × , {Pm }), one can define an EM (E, {Qm }) where the prob-
ability measures are given by QA,B,... (a, b, . . . ) = PA,B,... (a, b, . . . , λ).
λ∈Λ
Then, by definition, the given HVM is consistent with this “derived” EM.

19.2.2 Definitions of Properties of Hidden Variable Models

(HVM)

Strong Determinism (SD)

An HVM is Strongly Deterministic if the state of the system determines uniquely
the outcome of every measurement.
More formally, for every state λ the probability of observing each outcome for
every measurement is either 0 or 1: for every A ∈ M, λ ∈  ∃ a s. t. PA (a| λ) = 1.
Weak Determinism (WD)
An HVM is Weakly Deterministic if the state of the system determines uniquely the
outcomes of every admissible series of measurements.
Formally, for every state λ and for every admissible measurement series A, B, C,
. . . the probability of observing, a certain set of outcomes is either 0 or 1, i.e.

∃ (a, b, c, . . . ) s.t.PA,B,C,... (a, b, c, . . . |λ) = 1.

Note that in Strong Determinism the determinism applies to every measurement

individually, but in Weak Determinism it applies only to admissible sequences of
measurements. Hence SD implies WD but the converse is not true as can be seen
from the following examples.
Example 2.1 Assume a HVM with three measurements A, B, C each with a binary
outcome, two admissible measurement series (A, B), (A, C) and a two states λ = 0,
λ = 1 with equal probabilities.
19 Why a Relativistic Quantum Mechanical World Must Be Indeterministic 429

The HVM depicted by the following table

0, 0 0, 1 1, 0 1, 1
A, B |λ=0 1 0 0 0
A, C |λ=0 1 0 0 0
A, B |λ=1 0 0 0 1
A, C |λ=1 0 0 0 1

is Strongly Deterministic since the state λ = 0 implies a = b = c = 0 and λ = 1

implies a = b = c = 1.
Example 2.2 However, a similar setting with the table:

0, 0 0, 1 1, 0 1, 1
A, B |λ=0 0 1 0 0
A, C |λ=0 0 0 0 1
A, B |λ=1 0 0 1 0
A, C |λ=1 1 0 0 0

is Weakly Deterministic since given the state λ the outcome of the two admissible
measurements is uniquely determined, but it is not Strongly Deterministic since
given λ = 0 (and also λ = 1) the outcome of the measurement of A can be either
0 or 1 depending on whether it was measured together with B or with C. This
counter intuitive property which characterizes certain quantum systems is called
contextuality and will be investigated later in great details.
No-Signaling (Parameter Independence)3 (NS)
An HVM is No-Signaling if for every state λ, the outcome of a measurement does
not depend on the other measurements performed simultaneously with it.
Formally: for every λ: PA,B,... (a|λ) = PA,B  ,... (a|λ) = PA (a|λ).
The No-Signaling condition is associated with Special Relativity, since a HVM
that fails NS may allow superluminal signaling if the measurements are performed
in space-like separated locations.
Example 2.1 is No-Signaling but Example 2.2 is not as indicated before.

3 The term ‘Parameter Independence’ is more common in the literature, but due to our focus
on the relation between EM and SR we prefer the term ‘No-Signaling’, which emphasizes the
physical implication of this property. Also note that we use the term ‘signaling’ in a neutral
way, disregarding completely whether or not the signals can be noticed or readout, unless we say
otherwise. See Maudlin (1994) for a broader discussion of signaling.
430 A. Levy and M. Hemmo

Comment
Even when an HVM does not satisfy the No-Signaling property, it may still be
impossible to transmit superluminal signals. This is the case when it is not feasible4
to prepare the system at those states where violation of the NS property takes place.
Bohm’s (1952) theory is an example for this situation (see discussions of this point
and its implications in Pitowsky 1989; Dürr, Goldstein and Zanghi 1992; Albert
1992; Maudlin 1994; Hemmo and Shenker 2013).
Outcome Independence (OI)
An HVM is Outcome independent if for every state λ, the outcome of a measurement
does not depend on the outcomes of other measurements performed simultaneously
with it.
Formally: PA, B, C, . . . (a, b, c, . . . | λ) = PA, B, C, . . . (A = a| λ)PA, B, C, ...
(B = b| λ)PA, B, C, . . . (C = c| λ)· · · , for every λ.
Example 2.3 In tossing two independent binary coins with measurements A, B
) parameter λ ∈ {1, 2, 3, 4} characterized the 4 possible outcomes.
respectively a state
Let P (λ) = 1
4 for every λ then the HVM described by the following table
satisfied the Outcome Independence property:

0, 0 0, 1 1, 0 1, 1
A, B |λ=0 1 0 0 0
A, B |λ=1 0 1 0 0
A, B |λ=2 0 0 1 0
A, B |λ=3 0 0 0 1

Example 2.4 The following “two balls in two jars” scenario is an example of HVM
that is not Outcome-Independence.
A red ball and a blue ball are placed in two jars with equal probabilities. The
measurement of A detects the color of the ball in the first jar and B in the second jar.
Assume a singleton state space  and a probability table defined by:

)r) (b, b)
(r, r) (r,)b) (b,
(A, B) 0 1 1 0
2 2

It is clearly No-Signaling but not Outcome Independence since the outcomes

of the two measurements are dependent. Note that for the derived EM (which is

4 We don’t address here the question of why in a given theory the preparation of a system in a
certain state λ is not feasible.
19 Why a Relativistic Quantum Mechanical World Must Be Indeterministic 431

identical in this case to the HVM since  is a singleton) there is a consistent

HVM (with two states of equal probabilities) that is No-Signaling and Outcome
Independence which is given by

(r, r) (r, b) (b, r) (b, b)

(A, B) | λ = 0 0 1 0 0
(A, B) | λ = 1 0 0 1 0

Locality
The locality property is the conjunction of No-Signaling and Outcome Indepen-
dence properties, as defined by Bell (1964):
PA, B, C . . . (a, b, c, . . . | λ) = PA (a| λ)PB (b| λ)PC (c| λ)· · · for every λ.
λ – Independence (LI)
Brandenburger and Yanofsky (2008) define an additional property of HVM, which
they call λ-Independence (LI). This property entails that the state of the measured
system as determined by the value of the hidden variable is independent of the
measurement performed on the system and vice versa. However, in our settings of
HVM this definition is meaningless since we did not assume a probabilistic structure
on the admissible series of measurements. In fact, we tacitly assumed that there can
be no dependence between the set of admissible measurements and the state of
the measured system. Assuming the converse means that there is some sort of a
“universal conspiracy” that correlates between the choices of the experimenter and
the state of the measured system. For this reason, the λ-Independence property is
sometimes associated with lack of ‘free will’ (we don’t address this issue here).
As is well known, Bell (1964) has assumed λ-Independence. Although retro-
causal approaches to QM reject λ-Independence and admit correlations between the
choices of the experimenter and the state of the measured system (see e.g. Price and
Wharton 2013), we shall assume, as we said above, the more commonly accepted
assumption that λ-Independence holds.

19.2.3 Existence Theorems and Relations Between HVM

Properties

The following theorems and propositions are stated without their proofs. The
interested reader can find the proofs in Brandenburger and Yanofsky (2008) and
Brandenburger and Keisler (2016).
432 A. Levy and M. Hemmo

Existence Theorem
For every Empirical Model there exists a consistent Weakly Deterministic Hidden
Variable Model.
Note that this theorem does not guarantee that there is a consistent Strongly
Deterministic HVM. The most remarkable example for a Weakly Deterministic
HVM theory is Bohm’s interpretation of QM. As we shall see later it’s Weakly
Deterministic (but not SD) nature is intimately related to the fact that the values
assigned to all quantum mechanical observables other than position are contextual.
It follows from the above existence theorem that the Weakly Deterministic property
of Bohm’s theory is not an exception. However, it is still remarkable of course
that Bohm’s theory provides dynamics for the hidden variables that solves the
measurement problem in QM in a way that is compatible with the Schrödinger
dynamics of the quantum wavefunction, while this existence theorem refers only
to measurement outcomes without mentioning dynamics at all.
However, the generality of this theorem may come as a surprise since in the
history of QM it took a few decades to realize that von Neumann’s “no-go proof,”
according to which hidden variables theories consistent with the predictions of QM
do not exist, presupposed that the hidden variables have pre-measurement values
that can be measured as quantum observables (see Bub (2010) for a discussion of
this point). In the terminology of the EM and HVM framework, von Neumann’s
assumption is equivalent to requiring that the HVM satisfies Strong Determinism,
and indeed such an HVM does not exist for a contextual Empirical Model as we
shall explain later.
Comment on Strongly Deterministic Explanation
Brandenburger and Yanofsky (2008) consider another existence theorem:
Existence Theorem
For every Empirical Model there exists a consistent Strongly Deterministic HVM
(but not necessarily λ- independent).
Note that although this theorem guarantees that every Empirical Model has a
Strongly Deterministic HVM, this comes with the price that this model may not
satisfy λ – Independence (described at the end of Sect. 19.2.2). As we said, in
our setting we presuppose independence between the admissible measurements set
and the state of the measured system. Hence, in this setting λ – Independence is
guaranteed and therefore a Strongly Deterministic explanation may fail. We will see
in Sect. 19.4 that if the Empirical Model is classical (non-contextual; see below)
then it does have a consistent Strongly Deterministic HVM even in our setting (i.e.
an HVM that satisfies λ – Independence), but if it is contextual then a consistent
Strongly Deterministic HVM does not exist.
Relations Between HVM Properties (‘→’ Means Material Implication)
Let us state the following relations between properties of HVM without explicit
proofs:
19 Why a Relativistic Quantum Mechanical World Must Be Indeterministic 433

1. Strong Determinism → Weak Determinism → Outcome Independence

2. Strong Determinism → No signaling, hence
3. Strong Determinism → Locality (= Outcome Independence & No signaling)
4. Strong determinism ↔ Weak Determinism and No signaling, hence
5. Strong determinism ↔ Weak Determinism and Locality
Example 2.5) Consider two spin-half particles denoted 1, 2 that are in the singlet
state ψ = √2 (|↑1 |↓2 + |↓1 |↑2 ).
1


Let A, A denote the spin measurements of particle 1 in two different directions

and B, B the spin measurements of particle 2 in the same two directions. The
 
outcomes of the measurements are a ∈ {±1} for A, A and b ∈ {±1} for B, B .
Finally, let the state space be  = {0, 1} with uniform distribution and define the
measurement outcomes deterministically by the table below.
One can notice that this model is Weakly Deterministic since for every
(λ, m1 , m2 ) the tuple (a, b) is determined uniquely, but it is not Strongly Deter-

ministic as for (λ = 0, m1 = A, m2 = B), b = − 1 but for (λ = 0, m1 = A , m2 = B),
b = 1 and hence the outcome of measurement B is not determined by the state
alone. The model is Outcome Independent but is not No-Signaling which can be
verified directly or concluded from relations (1), (4) above, and therefore this model
is not local.

λ m1 m2 a b
0 A B 1 −1
1 A B −1 1
0 A B 1 1
1 A B −1 −1
0 A B 1 1
1 A B −1 −1
0 A B 1 −1
1 A B −1 1

19.3 Properties of Empirical Models and Their Relations

The properties of Outcome Independence (OI), No-Signaling (NS), Locality, Strong

Determinism (SD) and Weak Determinism (WD) for HVM have been defined and
analyzed in Brandenburger and Yanofsky (2008), Brandenburger and Keisler (2016)
and Abramsky (2013).
434 A. Levy and M. Hemmo

For EM however, No-Signaling is the only property defined (mistakenly as non-

contextuality) in Brandenburger and Yanofsky (2008), correctly in Abramsky (2013)
and more generally in Abramsky and Brandenburger (2011).5
An EM satisfies No-Signaling if the marginal distribution pertaining to a sub-
sequence of an admissible sequence of measurements does not depend on the rest
of the measurements in the sequence. More formally, for a sub-sequence of one
measurement:
 
For every admissible measurement series A, B, C, . . . and A, B , C , . . . , and
every a ∈ OA

QA,B,C,... (a) = QA,B  ,C  ,... (a) = QA (a).

We propose extending the definitions of all the HVM properties to Empirical

Models as follows. An Empirical Model has a certain property just in case there
exists a consistent HVM with this property. This extension is desirable since, as we
said in the Introduction, we wish to formulate the abovementioned properties in a
way that refers only to the phenomena and does not depend on a given theory.
Firstly let us show that our definition of No-Signaling is equivalent to the above
given definition:
By proposition 2.2 from Branderburger and Yanofsky (2008) if a HVM satisfies
No-Signaling and λ – independence then every equivalent EM satisfies No-Signa-
limg. In the other direction, if an EM is No-Signaling, then the trivial (i.e. with
singleton state space) equivalent HVM is No-Signaling by definition.
We now investigate the relations between the EM properties and the way they
differ from the HVM properties. Due to the existence theorem every Empirical
Model is either Strongly Deterministic or else it is Weakly Deterministic but not
Strongly Deterministic. Let us denote the second option (i.e. a WD model that is
not SD) by WD∗ . The significance of this distinction will be investigated in Sect.
19.4, which is devoted to the concept of contextuality. In contrast to the HVM
case, an EM that fails to satisfy No-Signaling allows for superluminal signaling (for
measurements performed at space-like separated locations) that can be empirically
carried out. Hence, to comply with Special Relativity an EM must be No-Signaling,
while it may have a consistent HVM that is not NS. We will show later that if an
Empirical Model is WD∗ it must have a consistent HVM that does not satisfy No-
Signaling.
Example 3.1 The EM derived from the HVM in Example 2.2 in Sect. (19.2)
satisfies No-Signaling, since the trivial HVM (singleton ) is NS by observing
thatPA, B (a) = PA, C (a) = PA (a) for a ∈ {0, 1}.

5 Inthis reference a notion of compatibility between marginal probability distributions is defined

as a generalization of the No-Signaling property which refers to one measurement marginal
probability only.
19 Why a Relativistic Quantum Mechanical World Must Be Indeterministic 435

a, b = 0, 0 0,)1 1,)0 1, 1
A, B 0 1 2 1 2 0
) )
A, C 1 2 0 0 1 2

However, the original HVM defined in this Example (2.2) is not No-Signaling
(since it is WD∗ ). This example demonstrates the fact that the phenomenal level EM
may have properties that are different from a specific physical HVM that realizes it.
By contrast to the HVM case, the conjunction of Outcome Independence and
No-Signaling for EM is not equivalent to the Locality property. This is due to the
fact that there may be a consistent HVM that is Outcome Independence and another
consistent HVM that is No-Signaling but there is no single consistent HVM that
satisfies both OI and NS. Example 3.1 above demonstrate this situation since it has
a Weakly Deterministic consistent HVM that is Outcome Independence (by relation
(1)) and the trivial HVM is No-Signaling but it is not local (i.e. does not have a local
HVM) since it is not classical. We will discuss this last point in Sect. 19.4.
Relations Between Properties of an Empirical Model
The first three relations pertaining to the properties of HVM properties (see Sect.
19.2) are satisfied also by EM:
1. Strong Determinism → Weak Determinism → Outcome Independence
2. Strong Determinism → No signaling, hence
3. Strong Determinism → Locality
However, relation (4): Weak Determinism and No signaling → Strong Deter-
minism, fails because (as we mentioned above) there can be a consistent HVM
which is WD and another consistent HVM which is No signaling but there is no
single consistent HVM with both properties. This situation is demonstrated by the
following example:
 
Example 3.2 Let there be four measurements A, A , B, B each with outcomes in
{0, 1} and the following conditional probability table:

(0,)0) (0, 1) (1, 0) (1,)1)

A B 1 2 1
) )2
A B 1 2 1
) )2
A B 1 2 1
) ) 2
A B  1 1
2 2
436 A. Levy and M. Hemmo

This EM is No-Signaling (the trivial HVM is NS) and it is Weakly Deterministic,

since every EM has a WD consistent HVM, but it is not Strongly Deterministic as
we explain in the next section.
Relation (5): Strong Determinism ↔ Weak Determinism and Locality, has in
fact a stronger version for Empirical Models. For an EM, Strong Determinism is
equivalent to Locality, a result that is proved in Abramsky and Brandenburger (2011,
Theorem 8.1). We discuss this result in Sect. 19.4.
Finally, note that if an HVM is No-Signaling then the derived (consistent) EM is
also No-Signaling. This is shown by Brandenburger and Yanofsky (2008).

19.4 Contextuality

Non-Contextuality is defined by Peres (1995) as the property that “the outcome

of experiments performed at a given location in space be independent of the
arbitrary choice of other experiments that can be performed, simultaneously, at
distant locations” (p. 187).
A common schematic example for contextuality (see e.g. Peres 1995) is the
following. Consider a measurement of an observable A (represented by a Hermitian
operator) that is performed simultaneously with a set of observables B1 , B2 , . . . ,
all of which pairwise commute with A. Then, consider another measurement of
A simultaneously with another set C1 , C2 , . . . that also commute pairwise with
A. Suppose that the operators B1 , B2 , . . . do not commute with C1 , C2 , . . . This
means that the outcome of the measurement of A may depend on the (measurement)
context in which it is performed.
The subtle point in this definition and example is that if ‘independence’ is
understood as “stochastic independence” (by which we mean that the distribution of
the probabilities of the outcomes depends on the measurement context), then contex-
tuality will allow superluminal signaling, which we assume here is impossible. This
point has been noted by, for example Badzia̧g et al. (2009) and this prompted them to
distinguish between what they call “contextuality of outcomes” and “contextuality
of probabilities.”
In the spirit of the above definition, Brandenburger and Yanofsky (2008) define
a non-contextual Empirical Model as one for which the probability of obtaining a
particular outcome of a measurement does not depend on the other measurements
performed. More formally:
 
An EM satisfies “non-contextuality” if for all a, A, B, B , C, C , . . . ,
QA,B,C,... (a) = QA,B  ,C  ,... (a) = QA (a).
Here we point out that this definition is identical to our definition of the No-
Signaling property for EM, given in Sect. 19.3 (it is also identical to the definition of
NS in Abramsky and Brandenburger 2011). Since No-Signaling must be satisfied for
every EM in order to be compatible with Special Relativity, using the contextuality
of probabilities property may not be a useful definition for differentiating “classical”
scenarios from quantum mechanical ones.
19 Why a Relativistic Quantum Mechanical World Must Be Indeterministic 437

For this reason, we propose that non-contextuality should mean something

similar to the idea of “non-contextuality of outcomes” (as in Badzia̧g et al. 2009).
Here the idea is that the outcome of the measurement of A should be independent
of which additional measurements, B1 , B2 , . . . or C1 , C2 , . . . , are performed
with it. However, this intuition cannot be directly captured by the EM framework
in which one only has probabilities for outcomes, not individual outcomes. This
intuition leads us to join Abramsky and Brandenburger (2011) and suggest that
non-contextuality for an Empirical Model is satisfied if there exists a consistent
“classical” HVM, which in our terminology means that it is Strongly Deterministic.
We argue that this is a useful definition since it sharply distinguishes between
a “classical” and a “quantum” EM (see below, for explicit definitions of these
notions), while No-Signaling is common to both.
Before presenting our definition of non-contextuality for EM we comment that
defining a similar property for HVM is superfluous since it will coincide with
previously defined terms. In the context of HVM, No-Signaling is equivalent to
non-contextuality of probabilities and Strong Determinism is equivalent to non-
contextuality of outcomes.
Characterization of a Classical Empirical Model
An EM is classical if its probability model is classical, i.e. it belongs to the convex
hull of the extreme probability distributions (those with zero or one probabilities
only).6
Therefore, it is possible to write:

QA,B,C,... (a, b, c, . . . ) = δ A,B,C,...,λ (a, b, c, . . . ) μ (λ) ,
λ

where δ A, B, C, . . . , λ are the extreme

 probability distribution indexed by a parameter
λ ∈  for some finite set  and μ (λ) = 1. Hence, the sum in the right-hand side
λ
of this formula is a convex combination and μ(λ) constitutes a probability measure
on the set of indices.
This characterization means that there exists an HVM, consistent with
the given EM, such that the state space of the hidden variables consists
precisely of the parameter set  and it is Strongly Deterministic. The reason
is that the δ A, B, C, . . . , λ are extremeprobabilities and hence can be written as
δ A,B,C,...,λ (a, b, c, . . . ) = δ X (x|λ), that is, every separate
x = a, b, c, . . .
X = A, B, C, . . .
measurement has exactly a single outcome that has probability 1.

6 SeePitowsky (1986) for characterizations of classical and quantum probability spaces as convex
subsets of the space of phenomenal distributions.
438 A. Levy and M. Hemmo

We now turn to the main definition of this section.

Definition: Non-Contextual Empirical Model An EM is non-contextual if and
only if it has a consistent Strongly Deterministic HVM, i.e.:

PA,B,C... (a, b, c, . . . , λ) = P (λ) δ X (x|λ) and
x = a, b, c, . . .
X = A, B, C, . . .
 
QA,B,C... (a, b, c, . . . ) = P (λ) δ X (x|λ)
λ
x = a, b, c, . . .
X = A, B, C, . . .
Now, in Abramsky and Brandenburger (2011, Theorem 8.1), and earlier in Fine
(1982) it is proved that an EM has a consistent Local (factorizable) HVM if and only
if it has a Strongly Deterministic HVM. Hence, we conclude that for an Empirical
Model the three properties: Strong Determinism, Locality and Non-Contextuality
are equivalent properties.
Therefore, there are only two possible scenarios for an Empirical Model:
It is either
• Classical (=Local = Strongly Deterministic = Non-Contextual), i.e., it has a
Strongly Deterministic consistent HVM.
Or else:
• It is Non-Classical (=Non-Local = Contextual), i.e., it has a Weakly Determin-
istic consistent HVM, but not a Strongly Deterministic one.
It seems to us that this result in a sense vindicates the arguments in Pitowsky
(2007) and Bub and Pitowsky (2010) that the essential character of quantum
mechanics is fixed by the non-classical, that is, the contextual nature of the quantum
probability.
Before concluding this section, we would like to mention the interesting question
of how to decide if a given EM is contextual or non-contextual. This question can be
formulated equivalently by asking if there exists a common “classical” probability
distribution Q for the outcomes of the entire set of measurements {A, B, C, . . . } such
that the family of probability distributions Qm defined each on one admissible series
of measurements are marginal distributions of Q.
It turns out that for a finite EM this question can be answered by constructing
a system of linear equations on the probability distributions that characterized the
model in such a way that it has a solution if and only if this EM is classical, namely
if it has a Strongly Deterministic HVM (see Abramsky and Brandenburger 2011).
Going into the details of this topic is beyond the scope of this paper, but let us only
say that this result allows us to show that, on the one hand, Examples 2.2, 2.5, 3.1
and 3.2 (the famous PR box model) are contextual, and hence do not have a Strongly
Deterministic HVM; on the other hand, Examples 2.1, 2.3, 2.4 are classical. Even
without resorting to this method we have been able to distinguish the two cases
by showing that the classical EM has a consistent Strongly Deterministic HVM
19 Why a Relativistic Quantum Mechanical World Must Be Indeterministic 439

and the contextual EM has a consistent Weakly Deterministic, but not Strongly
Deterministic HVM.
One last comment here which is important for the next section. If one wants
that a Contextual Empirical Model be compatible with Special Relativity, the model
must satisfy No-Signaling. Let there be a consistent Weakly Deterministic HVM
(which is guaranteed by the existence theorem). Then, this HVM cannot satisfy
No-Signaling, since if it did, it would be Strongly Deterministic (by property (4)
of Sect. 19.2), which implies that the Empirical Model is not contextual (contrary
to our assumption). We conclude that every deterministic HVM consistent with a
contextual EM must have a state that allows superluminal signaling! Therefore, in
order to comply with Special Relativity, it is either impossible to prepare a system in
such a state, or else a Weakly Deterministic HVM cannot exist. In the next section
we will present a strong argument in support of the second option.

19.5 Empirical Models and Special Relativity

The interplay between the non-locality of QM and the relativistic constraints

imposed by the Special theory of Relativity (SR) has drawn much attention in
the literature (see e.g. Redhead 1987 and references therein). A number of results
support the intuition that if one requires genuine adherence to Lorentz invariance,7
then even a nonlocal but deterministic explanation of certain quantum mechanical
scenarios leads to a contradiction (see e.g. Hardy and Squires 1992; Clifton et
al. 1992; Conway and Kochen 2006, 2009; Gisin 2011; Suarez 2010). Rohrlich
and Aharonov (2013, p. 206), for example, ask how can action at a distance be
compatible with relativistic causality, and their answer is that: “God plays dice,
because it is the only way for nonlocality and relativistic causality to coexist.”8
Similarly, Ben Menahem (2012, p. 164) argued that: “Once we distinguish between
nonlocality and no signaling, the lesson of QM is that a combination of nonlocality
and no signaling is made possible by indeterminism.” Here we shall generalize the
arguments from Gisin (2011) to the framework of Empirical Models and Hidden
Variable Models and consider the implications of this approach with respect to the
existence of Weakly Deterministic and Lorentz invariant HVM that is consistent
with a contextual EM.
Let us start by reiterating the roles played by EM and HVM as representations
of physical systems at different levels. An EM describes a physical system at the
phenomenological level, i.e., it depicts the distribution of outcomes of a given
set of measurements performed on the system. The HVM adds another layer to
the description, namely the physical state of the system. It may be impossible to

7 As we said in Sect. 19.1, by Lorentz invariance we mean that there is no preferred Lorentz frame
of reference and therefore no absolute time order and no absolute direction of causation.
8 This point has been stressed by Yakir Aharanov, since (as far as we know) the mid 1980s.
440 A. Levy and M. Hemmo

measure directly the value of this state, but when it is known, the probability distri-
bution over the measurement outcomes described by the EM can be explained. This
explanation may be deterministic (Strongly Deterministic or Weakly Deterministic
HVM) or otherwise stochastic. Therefore, the HVM plays the role of the physical
theory that allows making predictions about, or producing explanations for, the
results of measurements performed on the system.
Although the EM and HVM framework is aimed at analyzing physical systems,
it does not refer explicitly to space and time. Yet, the properties of No-Signaling
and Locality originated from physical systems that are positioned in space-time and
hence it is only natural to introduce these aspects to the EM and HVM framework.
A simple way of doing this is to assign to each measurement a space-time label
and then every pair of measurements can be in space-like, time-like or light-like
relation. And under these relations one can require, for example, that properties
such as No-Signaling or Outcome Independence will apply only to measurements
that are in the space-like relation. One can conclude (from special relativity) that
an EM that does not satisfy No-Signaling with respect to two space-like related
measurements cannot represent a physical system, since it will allow transmitting
information in superluminal speed.
A HVM, on the other hand (as explained in Sect. 19.2.2), may violate No-
Signaling but nevertheless it may constitute a deeper physical structure that gives
rise to the EM, since the HVM might also rule out the possibility of preparing a
system in a specific state, so that transmitting superluminal signals is not physically
feasible.
However, as we shall see below, requiring that a HVM is Lorentz Invariant in the
sense that it predicts consistent measurement outcomes in different inertial frames,
implies that the HVM must be No-Signaling, and from this fact we shall prove a no-
go theorem according to which a Weakly Deterministic HVM realizing a contextual
EM does not exist.
We stress that this no-go theorem is strictly stronger than the exclusion of a
strongly deterministic HVM for a contextual EM presented in Sect. 19.4, since
the existence theorem of Sect. 19.2.3, guarantees the existence of a Weakly
Deterministic HVM for every EM, including a contextual one. On the other hand,
our no-go theorem does not contradict this existence theorem, since it is the
Lorentz Invariance additional requirement that excludes the existence of the Weakly
Deterministic HVM. So, as we shall see below, our no-go theorem does not imply
that there is no Weakly Deterministic HVM consistent with a contextual EM. It
does, however, show that such a HVM will be inconsistent with the fact that the
time order of space-like measurements may be reversed for certain inertial frames.
To establish this no-go theorem let’s start with proving the following Lemma.
Lemma 5.1 An HVM that is consistent with Lorentz Invariance must be No-
Signaling.
By consistency with Lorentz Invariance we mean that given two measurements
in space-like locations the HVM yields identical probabilistic predictions for every
inertial frame of reference.
19 Why a Relativistic Quantum Mechanical World Must Be Indeterministic 441

Proof We generalize the argument in Gisin (2011), which is explicitly formulated

with respect to a Weakly Deterministic account of a Bell type scenario, to any
stochastic HVM pertaining to a measurement scenario with two space-like related
measurements.
Suppose that there are two observers Alice and Bob, in space-like locations, that
 
can perform measurements A, A and B, B , respectively, each with binary outcomes,
   
where all four combinations of measurements (A, B), (A , B), (A, B ), (A , B ) are
admissible series of measurements. Since the two measurement locations are in
space-like relation, there are two inertial Lorentz reference frames, such that in

one of them the measurement of A or A takes place before the measurement of

B or B , whereas in the other inertial frame the temporal order of the measurements
is reversed. We denote these two inertial frames by AB and BA, respectively, to
indicate the temporal order in which the measurements are performed (see Fig. 19.1
below).

Let us denote the outcomes of the measurements A, A by a, the outcomes of the

measurements B, B by b, and the parameter space of the HVM by .
   
For each admissible measurement (X, Y) ∈ {(A, B), (A , B), (A, B ), (A , B )} and
AB (X = a|λ) ,
each λ ∈  the HVM defines the conditional probabilities: PX,Y,...
P AB (Y = b|λ) and P BA (X = a|λ) , P BA (Y = b|λ). Now, in the AB
X,Y,... X,Y,... X,Y,...
AB (X = a|λ) cannot depend on Y, since the
Lorentz frame, the probability PX,Y,...

Fig. 19.1 Spacetime diagram of Bell type scenario

442 A. Levy and M. Hemmo


choice of measurement B or B may be taken after the measurement in Alice’s
AB (X = a|λ) = P AB (X = a|λ).
location.9 Hence PX,Y,... X
And similarly in the BA frame: P BA (Y = b|λ) = P BA (Y = b|λ).
X,Y,... Y
In addition, the two Lorentz frames must agree on the probability distributions
of every event and therefore one concludes that:

AB (X = a|λ) = P BA (X = a|λ) = P AB (X = a|λ)

PX,Y,... X,Y,... X

BA (Y = b|λ) = P AB (Y = b|λ) = P BA (Y = b|λ)

PX,Y,... X,Y,... Y
...

This implies that the HVM is No-Signaling.

Corollary A contextual EM has no Lorentz Invariant Weakly Deterministic con-
sistent HVM.
Proof Assume that such a consistent Weakly Deterministic HVM exists. Then,
since it is also No-Signaling (being Lorentz Invariant), it is in fact Strongly
Deterministic by relation (4) of Sect. 19.2.3. But this is a contradiction, since a
contextual EM cannot have a Strongly Deterministic consistent HVM.
Although, as we have seen, there is no deterministic (Weak or Strong) HVM of a
contextual EM, there always exists a probabilistic Lorentz Invariant and consistent
HVM, since the trivial consistent HVM has this property. So, in this sense our
analysis vindicates the above arguments (at the beginning of this section) that the
quantum mechanical nonlocal correlations are compatible with special relativity
only in a genuinely stochastic world.
It is important to mention that Abramsky and Brandenburger (2011) proved that
QM, which can be seen as a genuine stochastic HVM, with projective measurements
and quantum states playing the role of the λ parameter, is indeed No-Signaling!
If QM would not have this property it would be physically possible to transmit
superluminal signals, since quantum states can actually be prepared!
As we said, similar results have been derived in the literature for specific QM
scenarios.10 But, in contrast to these specific scenarios, the argument presented here
is general. It pertains to any contextual EM and does not depend on the details of
a specific theory. Even if QM would turn out to be incomplete or even wrong, the
framework of EM and HVM would remain intact. The proof above that there is
no deterministic HVM for contextual and Non-Signaling EM depends only on the
results obtained in experiments and on the requirement of Lorentz invariance of the
laws of physics.

9 We assume here λ-independence according to which there are no correlations between the choice
of measurement and the outcome of other measurements through the parameter λ.
10 See e.g., the two Mach-Zehnder type interferometers by Hardy and Squires (1992), the Kochen-

Specker set up by Conway and Kochen (2006, 2009), Bell’s EPR scenario by Gisin (2011), and
the Suarez-Scarai model by Suarez (2010).
19 Why a Relativistic Quantum Mechanical World Must Be Indeterministic 443

19.6 Application of Our No-Go Proof to Interpretations

of QM

Let us now briefly show how using the EM and HVM frameworks may be helpful in
dissolving certain debates that have been raised recently concerning the possibility
of constructing interpretations of QM that are Lorentz invariant. An argument
appearing in Conway and Kochen (2009) claims that for a contextual EM it is
possible to exclude also the existence of a Lorentz invariant consistent probabilistic
HVM. Conway and Kochen (2009) apply an analysis of the probability distribution
functions of a contextual EM which is similar to Gisin’s argument. However, their
conclusion is unsound, as indicated also in Goldstein et al. (2010) for a specific
scenario, since given a contextual, No-Signaling EM, the trivial (singleton )
consistent HVM is a No-Signaling Lorentz invariant probabilistic explanation for
this EM.
Our analysis shows that it is impossible to construct any deterministic Lorentz
invariant (and hence No-Signaling) HVM that yields the quantum mechanical
predictions. Since, for example, Bohm’s (1952) theory is Weakly Deterministic,
it must fail to be Lorentz invariant, and in this sense Bohm’s theory is incompatible
with Special Relativity (SR).11 The attempts to overcome this problem resort to
postulating that there is a preferred Lorentz reference frame which describes the
absolute temporal order of events (see Dürr et al. 2013). This absolute temporal
order circumvents our proof above exactly where we assume that a later event in
some reference frame cannot have an impact on an earlier event in the same frame.
The case of stochastic theories (of the kind originally proposed by Ghirardi et
al. 1986, GRW) is subtler, since prima facie a genuinely stochastic theory may
have a Lorentz invariant relativistic version compatible with QM, as we have shown
above. In Tumulka (2006) it is suggested that the flash theory denoted by rGRWf
is fully compatible with SR, in particular in the sense that it requires no preferred
Lorentz frame. Tumulka’s (2006) proposal has been criticized by Esfeld and Gisin
(2014) who claim that “neither rGRWf, nor any other theory, can account for the
occurrences of the Alice-flash and the occurrences of the Bob-flash in a Lorentz-
invariant manner”. This conclusion is based on the argument presented originally in
Gisin (2011) which we also used in this section to derive our no-deterministic HVM
result. Let us now examine Esfeld and Gisin’s (2014) argument in the context of the
rGRWf theory and the EM and HVM framework.
The core of Esfeld and Gisin’s (2014) argument (on our understanding) is that
the probabilities observed at space-like separation by Alice and Bob are “classical”,

11 For explicit proofs that Bohm’s theory requires an absolute temporal order (i.e., a preferred
Lorentz frame), see Albert (1992); Hemmo and Shenker (2013). Of course, as we said above,
Bohm’s theory does not allow to use its No-Signaling property, since its additional dynamical
law for the motion of particles together with its statistical postulate concerning the distribution of
particles entail that preparing a system at a known position that is more exact than the system’s
prepared wavefunction is physically impossible.
444 A. Levy and M. Hemmo

i.e. they can be “computed and simulated on a classical computer.” Hence, they
conclude that there exists a classical variable λ ∈  (that can be stored on a
computer’s hard disc) such that the outcomes on both sides of the EPR experiment
are given by the four deterministic functions that depend on this variable and on the
settings of the measurements as presented in our proof above. From this assumption,
they derive their conclusion according to which the rGRWf (and any other theory)
cannot provide a genuinely relativistic (i.e., Lorentz invariant) account of an EPR
experiment.
However, as we have just seen in this section, there always exists a probabilistic
HVM that is consistent with EPR-type experiments (or every other contextual
experiment) and this fact obviously disagrees with Esfeld and Gisin’s conclusion.
We will now argue that Esfeld and Gisin’s argument implicitly presupposes an
assumption which in fact need not hold in a stochastic theory, and specifically does
not hold in rGRWf.
Esfeld and Gisin (2014) are correct in arguing that the marginal probabilities of
an EPR empirical model, namely those computed by Alice or Bob, are classical.12
However, as explained in Sect. 19.4 above, this does not imply that there is a
common classical probability model for all possible combinations of measurements
on the two wings in the EPR setting. Assuming that the classical marginal
probabilities imply the existence of a common classical probability is equivalent
to assuming that the EM is non-contextual, which is not satisfied by the EM of the
EPR experiment.
So perhaps Esfeld and Gisin (2014) mean that there exists a consistent Weakly
Deterministic (non-classical) HVM for the EPR probabilities, which is indeed
true, as shown by Existence Theorem in Sect. 19.2. In this case one can derive a
contradiction with Lorentz invariance, as we have just done in this section. However,
even this assumption will not make their argument against the rGRWf theory valid,
unless one can prove that the rGRWf theory is Weakly Deterministic. But we do not
see how this can be done, because both the original GRW theory and the rGRWf
theory are manifestly genuinely stochastic theories. Of course, our point here should
not be taken as a positive proof that rGRWf is a genuinely relativistic interpretation
of QM. Such a proof seems to go beyond the framework of EM and HVM.

19.7 Conclusion

We saw in this paper that using the EM and HVM mathematical framework
allows us to clarify certain relations between significant physical properties such
as Determinism, Locality and Contextuality. Moreover, we were able to have an
insight into the debate concerning the prospects of interpretations of QM that are

12 A non-classical EM may have many classical marginal probabilities; this is the core of Abramsky

and Brandenburger (2011).

19 Why a Relativistic Quantum Mechanical World Must Be Indeterministic 445

compatible with Special Relativity. In particular we have proved here using the EM
and HVM framework that the physical reality of our world (assuming that there is
no absolute time order) must be stochastic.13
Viewed from this framework one can even think of standard QM (in von
Neumann’s formulation) as a stochastic HVM (in which the quantum state plays
the role of the hidden variable) that is consistent with the observed frequencies of
measurement outcomes. As we saw in Sect. 19.5, from a mathematical point of view,
Lorentz invariance (in the sense that there is no preferred Lorentz frame)14 by itself
is consistent with a genuinely stochastic HVM that yields the predictions of QM.
The genuine stochastic nature of QM allows it to have the property of No-Signaling
even at the HVM level.
However, the EM and HVM framework cannot really settle the question of
the compatibility of QM and SR because the dynamical aspect of physics is
missing in this framework. And obviously, one should not expect to find in this
framework a resolution to the measurement problem in QM. But it is nonetheless
encouraging that there is no mathematical barrier to finding a stochastic theory that
is compatible with both QM and SR. Whether or not rGRWf or other attempts
should be considered satisfactory from a physical point of view is still to be settled.

Acknowledgement We thank Orly Shenker and Yemima Ben-Menahem for discussions of issues
related to this paper. We also thank Ori Feiglin, Aharon Gill, Zohar Maliniak and Moti Suess, from
the research group in philosophy of physics at the University of Haifa for valuable comments on
this paper. Finally, we thank Samson Abramsky for very valuable comments on an earlier draft of
this paper. This research has been supported by the Israel Science Foundation (ISF), grant number
1148/18.

References

Abramsky, S. (2013). Relational hidden variables and non-locality. Studia Logica, 101(2), 411–
452.
Abramsky, S., & Brandenburger, A. (2011). The sheaf-theoretic structure of non-locality and
contextuality. New Journal of Physics, 13, 113036.
Albert, D. (1992). Quantum mechanics and experience. Cambridge, MA: Harvard University Press.
Badzia̧g, P., Bengtsson, I., Cabello, A., & Pitowsky, I. (2009). Universality of state-independent
violation of correlation inequalities for noncontextual theories. Physical Review Letters, 103,
050401.
Bell, J. (1964). On the Einstein-Podolsky-Rosen paradox. Physics, 1, 195–200.
Ben-Menahem, Y. (2012). Locality and determinism: The odd couple. In Y. Ben-Menahem & M.
Hemmo (Eds.), Probability in physics (pp. 149–166). Berlin/Heidelberg: Springer.

13 We did not go into the question why God’s dice is quantum rather than, say, a PR dice. This
will involve adding extra assumptions that inhibit the appearance of super-quantum theories (see
Rohrlich and Aharonov 2013 for a recent survey).
14 Of course, in a relativistic quantum theory, the Schrödinger dynamics is replaced with a

relativistic equation of motion.

446 A. Levy and M. Hemmo

Bohm, D. (1952). A suggested interpretation of the quantum theory in terms of “hidden” variables.
Physical Review, 85(166–179), 180–193.
Brandenburger, A., & Keisler, H. J. (2016). Fiber products of measures and quantum foundation.
In J. Chubb, A. Eskandarian, & V. Harizanov (Eds.), Logic and algebraic structures in quantum
computing, lecture notes in logic (Vol. 45, pp. 71–87). Cambridge: Cambridge University Press.
Brandenburger, A., & Yanofsky, N. (2008). A classification of hidden variable properties. Journal
of Physics A: Mathematical and Theoretical, 41, 425302.
Bub, J. (2010). Von Neumann’s ‘no hidden variables’ proof: A re-appraisal. Foundations of
Physics, 40(9), 1333–1340.
Bub, J., & Pitowsky, I. (2010). Two dogmas about quantum mechanics. In S. Saunders, J. Barrett,
A. Kent, & D. Wallace (Eds.), Many worlds? Everett, quantum theory and reality (pp. 433–
459). Oxford: Oxford University Press.
Clifton, R., Pagonis, C., & Pitowsky, I. (1992). Relativity, quantum mechanics and EPR. In D. Hull,
M. Forbes, & K. Okruhlik (Eds.), PSA 1992 (Vol. 1, pp. 114–128). East Lansing: Philosophy
of Science Association.
Conway, J., & Kochen, S. (2006). The free will theorem. Foundations of Physics, 36(10), 1441–
1473.
Conway, J., & Kochen, S. (2009). The strong free will theorem. Notices of the American
Mathematical Society, 56(2), 226–232.
Dürr, D., Goldstein, S., & Zanghi, N. (1992). Quantum equilibrium and the origin of absolute
uncertainty. Journal of Statistical Physics, 67(5/6), 843–907.
Dürr, D., Goldstein, S., & Norsen, T. (2013). Can Bohmian mechanics be made relativistic?
Proceedings of the Royal Society A, 470, No. 2162, 20130699.
Esfeld, M., & Gisin, N. (2014). The GRW flash theory: A relativistic quantum ontology of matter
in space-time? Philosophy of Science, 81(2), 248–264.
Fine, A. (1982). Hidden variable, joint probability and the Bell inequalities. Physical Review
Letters, 48(5), 291.
Ghirardi, G., Rimini, A., & Weber, T. (1986). Unified dynamics for microscopic and macroscopic
systems. Physical Review, D, 34, 470–479.
Gisin, N. (2011). Impossibility of covariant deterministic nonlocal hidden-variable extensions of
quantum theory. Physical Review A, 020102(R), 83.
Goldstein, S., Tausk, D. V., Tumulka, R., & Zanghi, N. (2010). What does the free will theorem
actually prove? Notices of the American Mathematical Society, 57, 1451–1453.
Hardy, L., & Squires, E. J. (1992). On the violation of Lorentz-invariance in deterministic hidden-
variable interpretations of quantum theory. Physics Letters A, 168(3), 169–173.
Hemmo, M., & Shenker, O. (2013). Probability zero in Bohm’s theory. Philosophy of Science, 80,
1148–1158.
Kolmogorov, A. N. (1933). Foundations of the theory of probability. New York: Chelsea Publishing
Company. English translation 1956.
Maudlin, T. (1994). Quantum non-locality and relativity: Metaphysical intimations of modern
physics. Oxford: Wiley-Blackwell Press.
Peres, A. (1995). Quantum theory: Concepts and methods. Dordrecht: Kluwer Academic Press.
Pitowsky, I. (1986). The range of quantum probability. Journal of Mathematical Physics, 27, 1556.
Pitowsky, I. (1989). Quantum probability, quantum logic. University of California, Springer-
Verlag.
Pitowsky, I. (2007). Quantum mechanics as a theory of probability. In: W. Demopoulos & I.
Pitowsky (Eds.), Physical theory and its interpretation, essays in honor of Jeffrey Bub, the
Western Ontario series in philosophy of science (Vol. 72, pp. 213–240). Springer.
Popescu, S., & Rohrlich, D. (1994). Quantum nonlocality as an axiom. Foundations of Physics,
24(3), 379–385.
Price, H., & Wharton, K (2013). Dispelling the quantum spooks – a clue that Einstein missed?
arXiv 1307.7744 [physics.hist-ph].
Redhead, M. (1987). Incompleteness, nonlocality and realism. Oxford: Clarendon.
19 Why a Relativistic Quantum Mechanical World Must Be Indeterministic 447

Rohrlich, D., & Aharonov, Y. (2013). PR-box correlations have no classical limit. In D. C.
Struppa & J. M. Tollaksen (Eds.), Quantum theory: A two-time success story (Yakir Aharonov
festschrift) (pp. 205–211). New York: Springer.
Suarez, A. (2010). The general free will theorem. arXiv: 1006.2485 [quant-ph].
Tumulka, R. (2006). A relativistic version of the Ghirardi–Rimini–Weber model. Journal of
Statistical Physics, 125, 821–840.
Chapter 20
Subjectivists About Quantum
Probabilities Should Be Realists About
Quantum States

Wayne C. Myrvold

Abstract There is a significant body of literature, which includes Itamar Pitowksy’s

“Betting on the outcomes of measurements,” that sheds light on the structure of
quantum mechanics, and the ways in which it differs from classical mechanics, by
casting the theory in terms of agents’ bets on the outcomes of experiments. Though
this approach, by itself, is neutral as to the ontological status of quantum observables
and quantum states, some, notably those who adopt the label “QBism” for their
views, take this approach as providing incentive to conclude that quantum states
represent nothing in physical reality, but, rather, merely encode an agent’s beliefs.
In this chapter, I will argue that the arguments for realism about quantum states go
through when the probabilities involved are taken to be subjective, if the conclusion
is about the agent’s beliefs: an agent whose credences conform to quantum
probabilities should believe that preparation procedures with which she associates
distinct pure quantum states produce distinct states of reality. The conclusion can be
avoided only by stipulation of limitations on the agent’s theorizing about the world,
limitations that are not warranted by the empirical success of quantum mechanics or
any other empirical considerations. Subjectivists about quantum probabilities should
be realists about quantum states.

Keywords Quantum mechanics · Quantum ontology · Pusey-Barrett-Rudolph

theorem · QBism · Subjective probability

20.1 Introduction

There is a conception of probability, which often goes by the name of subjective

probability, that takes probability to have to do with an agent’s degrees of belief.
An important step in the development of this conception was its integration with

W. C. Myrvold ()
Department of Philosophy, The University of Western Ontario, London, ON, Canada
e-mail: wmyrvold@uwo.ca

© Springer Nature Switzerland AG 2020 449

M. Hemmo, O. Shenker (eds.), Quantum, Probability, Logic,
Jerusalem Studies in Philosophy and History of Science,
https://doi.org/10.1007/978-3-030-34316-3_20
450 W. C. Myrvold

decision theory, pioneered by Ramsey and de Finetti, and developed more fully
by Savage. On this view, an agent’s preferences between acts (often illustrated by
choices of bets to make or accept) are taken to be indications of the agent’s degrees
of belief in various propositions.
It is an attractive idea to extend this conception to quantum probabilities. One
systematic development of this idea can be found in Pitowsky’s “Betting on the
outcomes of measurements: a Bayesian theory of quantum probability” (Pitowsky
2003). The idea also forms the basis of the position (or family of positions) known as
QBism (Caves et al. 2002, 2007; Fuchs and Schack 2013; Fuchs et al. 2014; Fuchs
2017; Fuchs and Stacey 2019). Proponents of such views tend to also take quantum
states to be nothing more than codifications of an agent’s beliefs. But the argument
for this conclusion is never spelled out with sufficient clarity.
On the other hand, we find arguments—of which the theorem of Pusey et al.
(2012) is the best known—for the opposite conclusion, namely, that quantum
states should be taken to represent elements of physical reality. If interpreting
quantum probabilities as subjective leads one inexorably to rejecting an ontic view
of quantum states, then there must be something about the reasoning behind these
arguments that doesn’t go through on a subjective reading of probability. In this
chapter, I will argue that this is not the case. Suitably interpreted, a version of the
PBR argument leads to the conclusion that an agent who takes quantum mechanics
as a guide for setting her subjective degrees of belief about the outcomes of future
experiments ought to believe that preparations with which she associates distinct
pure states result in ontically distinct physical states of affairs (this terminology will
be explained in Sect. 20.3, below). Note that this conclusion is entirely about the
agent’s beliefs. We are not presuming, as part of the argument, that the agent’s
beliefs in any way reflect physical reality. That is a matter for the agent’s own
judgment (though, of course, there is an implication that if you take quantum
mechanics as a guide for setting your degrees of belief about the outcomes of future
experiments, you should take distinct pure states to be ontically distinct).
The argument, of course, involves assumptions about the agent’s credences. It is
also assumed that the agent may entertain theories about the way the world is, and
have degrees of belief in such theories. It is, of course, logically possible to reject
all such theorizing. If such a move is proposed in connection with quantum theory,
then the question arises whether the empirical evidence that led the community of
physicists to accept quantum theory over classical theory provides any incentive
for the move. I will argue that we have no reason whatsoever to make such a
move. Anyone is, of course, free to make such a move without reason, as a free
choice. Dissuasion of someone so inclined will not be the topic of this chapter; what
concerns us here is the evidential situation.

20.2 Credence and Action: Some Preliminary Considerations

In this section I would like to bring to the reader’s mind some considerations of
the sort that may be called “pre-theoretical”—the sort of considerations that tend
20 Quantum Probabilities 451

to be taken for granted, prior to any physical theorizing. They are, for that reason, to
be thought of as neither particularly classical nor quantum. I do not take myself to
be saying anything new, radical, or contentious. Of course, one may start a scientific
investigation with certain presuppositions about the way the world is, and end up
concluding, as a result of one’s investigations, that some of those presuppositions
are false. If any readers are inclined to dispute some assertion made in this section,
then I will ask that reader to provide evidence for the falsity of the assertion. I
hope that it goes without saying that the mere fact that some proposition leads to an
unwelcome conclusion does not suffice as evidence that the proposition is false.
Consider the following situation. You are attending a lecture, and you are taking
notes on a laptop. This process involves a sequence of choices: you are choosing, at
every moment, which, if any, keys to hit. Why are you doing this? If you are taking
notes for your own benefit, as an aid to memory, then, presumably, you believe that,
at some later date, you will be able to see or hear or feel (depending on the device
used for readout) something that will be informative about your prior choices of
key-strokes. If you are taking notes for someone else’s benefit, then, presumably,
you believe that, at a later date, that other agent will be able to see or hear or feel
something that will be informative about your earlier choices of key-strokes.1
Why are you using a laptop, rather than, say, a kumquat? Presumably, you believe
that there is something about the internal structure of the laptop that makes it suitable
for the purpose. Suppose that you can use the readout to distinguish between two
alternative choices of things to write today. For concreteness, consider a choice
between writing
A ‘Twas brillig, and the slithy toves did gyre and gimble in the wabe.
B Now is the winter of our discontent made glorious summer by this sun of York.
If you believe that the readout tomorrow will allow you to distinguish between
having made choice A and choice B today (in case you’re forgetful)—that is, if
you believe that the laptop can serve as an auxiliary memory—this is to attribute
to the laptop a certain role in a causal process linking your choice today to your
experience tomorrow. Even if you have little or no idea of the internal workings of
the laptop, attribution of this causal role to the laptop involves an assumption that
your action will have an effect on the internal state of the laptop, and that the effect
your action had on the internal state of the laptop will, in turn, have an effect on
what it is that you experience tomorrow.
Now, an extreme version of operationalism might forbid you to entertain
conjectures about the internal state of the laptop, and speak only of the relations
between your act today and the readout tomorrow, with no mention of intervening
variables. But the fact is that you are using a laptop rather than a kumquat, and,
in fact, would not attempt to use a kumquat for this purpose, and that means that

1 One might also consider the following situation. You are reading a text of which you are not the
author. Presumably, you take what you are reading as informative about what choices of keystrokes
the author made while composing it. If not, then why are you still reading?
452 W. C. Myrvold

you believe that there is something about the physical constitution of the laptop that
makes it suitable for the role you wish it to play in the causal chain between your act
today and the readout tomorrow. Caution should be exercised in theorizing about the
internal workings of the laptop, but, as long as we are careful not to be committed
to anything more than what we have good evidence for, there is no harm, and some
potential benefit, in theorizing about the processes that mediate your actions today
and experiences tomorrow.
Suppose, then, that you begin—tentatively, and with all due caution—to form
a theory about the inner workings of the laptop. What constraints do your beliefs
about the input-output relations of the laptop place on the sorts of theories you
should entertain about its inner workings?
It seems that there are a few general things that can be said. If you believe
that the laptop can serve as a means for discriminating (with certainty) between
act A and act B, you should believe that there are distinct sets of internal states
corresponding to these two acts. That is, the set of states that the laptop could end
up in, as a result of your performing act A, has no overlap with the set of states that
it could end up in as a result of your performing act B. Moreover, these two sets of
states are distinguishable upon readout—the readout mechanism produces outputs,
corresponding to the two sets of internal states associated with acts A and B, that
are perceptually distinguishable by you. The only alternative to believing this is to
believe in a direct causal effect of your choice today on your experience tomorrow,
unmediated by any effect on the state of the world in the interim.
Of course, discrimination with absolute certainty is too much to ask. You might
not regard your laptop as a completely reliable means for obtaining information
about your past choices of key-strokes. That’s okay; it can still be informative, as
long as if you take some perceived outputs to be more likely, given some choices of
keystrokes, than others. We can cash this out in terms of your conditional degrees
of belief.
All of this will be formalized in the next section.

20.3 Constructing a Framework

Let us begin to formalize the informal considerations of the previous section.2

To simplify matters, we will imagine a Bayesian agent, Alice, whose gradations
of strength of belief in various propositions can be represented by a real-valued
function satisfying the axioms of probability. We will call this function the agent’s
credence function, and the value that it takes on a given proposition p, the agent’s
credence in p. We will avoid talking of “probability,” in case that, for some, the
word has connotations of something non-subjective.

2 Theframework in this section is based on the ontological models framework of Harrigan and
Spekkens (2010).
20 Quantum Probabilities 453

We consider some physical system (you may think of the laptop as a running
example), with which our agent will engage. She has some set of operations
{φ 1 , . . . , φ n } that can be performed on the system at time t0 (think choices of key-
strokes). We will also consider situations in which an act is chosen without Alice’s
knowledge of which one; she has an n-sided die that she can toss (and when she
does, her credence is equally divided among the possible outcomes), and a gadget
that will choose among the operations {φ 1 , . . . , φ n }, depending on the outcome of
the die-throw.
At some later time t1 , she performs an operation R (a “readout” operation), which
will result in one of a set of perceptually distinguishable results {r1 , . . . rm }. (We
can, of course, generalize, and give her a choice of operations to perform, but for
our purposes, one will suffice.)
We assume that our agent has conditional credences, cr A (rk |φ i ), representing
credence in result rk on the supposition of act φ i . Upon learning the result of the
operation R, she uses these to update, via conditionalization, her credences about
which act was performed.
Alice does not believe that the acts {φ i } have a direct influence on the outcomes
{rk }, unmediated by any effect of those acts on the system S. We will permit her
to entertain various theories about the workings of the system. Any theory T of
that sort will involve a set ΩT of possible physical states. At the level of generality
at which we are operating, we will assume nothing about the structure of these
state-spaces ΩT , and, in particular, do not assume that they are either classical or
quantum in structure. We do assume that, on the supposition of theory T , Alice can
form credences in propositions about the state of the system S to the effect that the
system’s state is in a subset Δ of ΩT , and that, for each theory T , there is an algebra
AT of subsets of ΩT deemed suitable for credences of that sort.
Suppose, now that Alice regards the state of the system to be relevant to the
outcome of the readout operation. What she thinks about that relevance may, of
course, depend on the theory T . To represent these judgments, we suppose her to
have conditional credences of the form cr A (rk |φ i , T , ω). This is her conditional
credence, on the supposition of theory T and the supposition that the state of
the system is ω ∈ ΩT , that the result of the readout will be rk , if act φ i was
performed. It should be stressed that these are Alice’s conditional credences; there
is no assumption that these conditional credences mirror any causal structure out
there in the rest of the world. Alice may hope that her conditional credences reflect
something of the causal structure of the world (or rather, the causal structure,
according to theory T ), but no assumption to the effect that they do reflect the causal
structure of the world will form any part of our argument, because our conclusions
will only be about what Alice should believe.
We assume, also, that, on the supposition of T , Alice has credences about
the states produced by the acts φ i . For any subset Δ of ΩT that is in AT , let
cr A (Δ|T , φ i ) be Alice’s conditional credence, on the supposition of theory T , that
performing act φ i would put the system into a state in Δ. Her credences about the
state should mesh with her conditional credences cr A (rk |φ i , T , ω) to yield her act-
outcome conditional credences, on the supposition of theory T :
454 W. C. Myrvold

cr A (rk |φ i , T ) = cr A (rk |φ i , T , ω), (20.1)

where the angle-brackets  ·  indicate expectation value taken with respect to her
credences cr A (Δ|T , φ i ) about the state, given act φ i . If {Tj } is the set of theories
that Alice takes seriously enough to endow with nonzero credence, then her act-
outcome conditional credences cr A (rk |φ i ) should satisfy

cr A (rk |φ i ) = cr A (rk |φ i , Tj ) cr A (Tj ). (20.2)
j

As mentioned, Alice does not believe that the acts {φ i } have a direct influence on
the outcomes {rk }, unmediated by any effect of those acts on the system S. This will
be reflected in her conditional credences; her belief can be captured by the condition
that cr A (rk |φ i , T , ω) be independent of which act is performed. That is, for all i, j ,

cr A (rk |φ i , T , ω) = cr A (rk |φ j , T , ω), (20.3)

for all states ω ∈ ΩT . This is the condition that her choice of act has an influence
on the later outcome only via the influence of this choice on the state of the system.
Now, part of the point of taking notes on a laptop is that what you see later will
permit you to distinguish between choices of key-strokes you made earlier. There
may be a limit to such discrimination; if you type something and then delete it,
the later readout might fail to distinguish between alternatives for the deleted text.
Some pairs of acts will be readout-distinguishable; others will not. We will say that
Alice takes two acts {φ i , φ j } to be R-distinguishable if, for every outcome rk of the
operation R, either cr A (rk |φ i ) or cr A (rk |φ j ) is zero.
If Alice takes two acts to be R-distinguishable, then, provided she takes the
influence of acts performed on the system on the later readout to be mediated by
a change in the state of the system, then she should also take it that the sets of states
that can result from the two acts are distinct. We can formalize this; say that Alice
takes the effects of acts {φ i , φ j } to be ontically distinct, on the supposition of theory
T , if there exists Δ ⊆ AT such that cr A (Δ|T , φ i ) = 1 and cr A (Δ|T , φ j ) = 0. We
have the following simple theorem.
Theorem 1 If an agent’s credences satisfy (20.1), (20.2) and (20.3), then, if she
takes two acts to be R-distinguishable for some operation R, then she also takes
the effects of those acts to be ontically distinct on any theory T in which she places
nonzero credence.
We emphasize: this is a theorem about Alice’s credences, not about actual states of
the system. Alice could be wrong in her judgment that the effects on her laptop of
different choices of keystrokes are ontically distinct. It might be that her keystrokes
have no effect whatsoever on the internal state of the laptop (in which case she
would be misguided in taking those acts to be readout-distinguishable).
All of this is very general, and no assumptions have been made about the
sorts of theories that Alice might entertain. Now, suppose we apply it to quantum
20 Quantum Probabilities 455

mechanics, with quantum probabilities construed subjectively. Suppose that Alice’s

credences about the results of future readout operations performed on system S can
be represented by quantum states, and that which quantum state it is that represents
her credences depends on her choice of act. If two acts are distinguishable by
some experiment that can be done on the system, the corresponding quantum states
are orthogonal. If Alice places nonzero credences only in theories for which her
conditional credences satisfy (20.3), then she should also believe that the effects of
acts with which she associates orthogonal quantum states are ontically distinct.

20.4 Ontic Distinctness of Non-orthogonal States

We have come to the conclusion that Alice should take the effects of preparation-acts
that she regards as R-distinguishable for some operation R to be ontically distinct.
This includes acts with which she associates orthogonal quantum states. But what
about non-orthogonal quantum states?
Let {φ 1 , . . . , φ n } be a set of preparation-acts that are such that Alice’s credences
about the results of future experiments, upon performance of these acts, can be
represented by pure quantum states {|φ 1 , . . . , |φ n }. If Alice takes the effects of
these acts to be ontically distinct, this means that she believes that the physical
state carries a trace of which preparation was performed and, that, if she knew
the physical state, this would be enough to specify which of the quantum states
{|φ 1 , . . . , |φ n } she deems appropriate to use to set her credences about the
results of future experiments. That is, she believes that, for each physical state that
could arise from one of these preparations, there is a unique quantum state that
“corresponds” to it as the state to use in setting her credences. The correspondence
might not be one-one, as there might be a multiplicity of physical states that
correspond to the same quantum state.
If Alice believes that the effects of any pair of preparation-acts with which she
associates distinct pure quantum states are ontically distinct, then Alice believes that
there is something in physical reality corresponding to these quantum states. We will
say, in such a case, that Alice is a realist about pure quantum states.
The PBR theorem can be adapted to provide a set of conditions on Alice’s
credences sufficient for her to be a realist about pure quantum states.
Following the terminology of Leifer (2014), we will say that Alice takes a set
{φ 1 , . . . , φ n } of preparation-acts to be antidistinguishable if there is an operation R
such that, for each result rk , cr A (rk |φ i ) is zero for at least one φ i . Though Alice
might not be able to uniquely decide which of the acts was performed, she will
always be able to rule at least one out, whatever the result of operation R is.
The setup of the PBR theorem is as follows. Suppose that we have two systems,
S1 and S2 , of the same type, and suppose that each of them can be subjected to
either of two preparation-acts, {φ, ψ}. We assume that each choice of preparation
act performed on one system is compatible with both choices of act performed on
456 W. C. Myrvold

the other. This gives us four possibilities of preparation-acts for the joint system,
which we will designate by {φ 1 ⊗ φ 2 , φ 1 ⊗ ψ 2 , ψ 1 ⊗ φ 2 , ψ 1 ⊗ ψ 2 }.
Suppose, now, that Alice regards these four preparation-acts as antidistinguish-
able. Under certain further conditions, which we will now specify, she should then
take the effects of the preparation-acts φ and ψ on the two systems to be ontically
distinct.
A theory about the physical states of the two systems, presumably, would be able
to treat S1 and S2 on their own, and also to regard the composite system that has S1
and S2 as parts as a system in its own right. It will, therefore, include state-spaces
Ω1 and Ω2 for the individual systems, and a state space Ω12 for the composite
system. In a classical theory, to specify the physical state of the composite system
S1 + S2 , it suffices to specify the states of the component parts, and so the state
of the composite can be represented as an ordered pair ω1 , ω2  of states of the
components, and Ω12 can be taken to be the set of all such ordered pairs. That is,
Ω12 can be taken to be the cartesian product of Ω1 and Ω2 . For other theories, such
as quantum mechanical theories, the cartesian product Ω1 × Ω2 might be included
in Ω12 without exhausting it. In quantum mechanics, for any states ψ 1 , ψ 2 of two
disjoint systems, there is a corresponding product state ψ 1 ⊗ ψ 2 of the composite,
but not all pure states of the composite system are of this form, as there also pure
entangled states.
We will say that a theory T satisfies the Cartesian Product Assumption (CPA)
if Ω1 × Ω2 is a subset of Ω12 . Given a theory satisfying the CPA, Alice’s
credences regarding a pair of preparations φ 1 , φ 2 that can be performed on the
two subsystems will be said to satisfy the No Correlations Assumption (NCA) if
her credences about the joint state of the two systems, on the supposition of T
and the preparations φ 1 , φ 2 , are concentrated on the cartesian product Ω1 × Ω2
and if her credences are such that information about the state of one system would
be completely uninformative about the state of the other. The conjunction of the
Cartesian Product Assumption and the No Correlations Assumption is called the
Preparation Independence Postulate (PIP). Note that this is a combination of an
assumption about the structure of the state space of a theory T and an assumption
about Alice’s credences, conditional on the supposition of T and on the preparations
φ1, φ2.
Now suppose that we have a pair of systems S1 , S2 , a theory T about them that
satisfies the CPA, and, for each system, a pair of preparation-acts φ i , ψ i , which are
such that:
1. There is an operation R that can be performed on the pair of systems, such that,
for each outcome rk of the operation, at least one of

{cr A (rk |φ 1 ⊗ φ 2 ), cr A (rk |φ 1 ⊗ ψ 2 ), cr A (rk |ψ 1 ⊗ φ 2 ), cr A (rk |ψ 1 ⊗ ψ 2 )}

is equal to zero. That is, Alice regards these preparation-acts as antidistinguish-

able.
2. Alice’s credences about the states of the systems S1 and S2 , conditional on the
supposition of theory T and each of the preparations mentioned, satisfy the NCA.
20 Quantum Probabilities 457

It then follows, by the argument of Pusey et al. (2012), that Alice should regard the
effects of these preparations to be ontically distinct, conditional on the supposition
of theory T .
Now, of course, Alice might entertain theories for which the CPA does not hold,
and, among those for which it does, her conditional credences might satisfy the
NCA for some but not for others. In this case, she should not be certain that the
effects of these preparations are ontically distinct. However, her credence that they
are must be at least as high as her total credence in the class of theories for which
the above-listed conditions are true. Another way to put this is: in order to be certain
that the effects of these preparations are not ontically distinct—that is, in order to
categorically deny onticity of quantum states—Alice’s credences in all such theories
must be strictly zero.
Though it seems natural from a classical standpoint, quantum mechanics gives us
incentive to question the Cartesian Product Assumption. In the set of pure states of
a bipartite system, product states are rather special, and every neighbourhood of any
product state contains infinitely many entangled states. If both systems have Hilbert
spaces of infinite dimension, then the entangled states are dense in the set of all
states, pure or mixed (see Clifton and Halvorson 2000, 2001). We should, therefore,
take seriously the cases of theories whose state spaces either lack a Cartesian product
component, or are such that product states cannot reliably be prepared exactly,
although we may be able to approximate them arbitrarily closely. A modification
of the PBR argument that does not employ the Cartesian Product Assumption is
therefore desirable.
In Myrvold (2018, 2020), a substitute for the PIP is proposed, which is strictly
weaker than it and dispenses with the Cartesian Product Assumption. This substitute
is called the Preparation Uninformativeness Condition (PUC).
To state the assumption, we consider the following set-up. Suppose that, for
systems S1 , S2 , we have some set of possible preparation-acts that can be performed
on the individual systems. Suppose that the choice of preparation for each of the
subsystems is made independently, say, by rolling two separate dice. Following
the preparation of the joint system, which consists of individual preparations on
the subsystems, you are not told which preparations have been performed, but you
are given a specification of the ontic state of the joint system. On the basis of this
information, you form credences about which preparations were performed. In the
case of preparations whose effects you take to be ontically distinct, you will be
certain about what preparation has been performed; otherwise, you may have less
than total information about which preparations were performed.
We ask: under these conditions, if you are now given information about which
preparation was performed on one system, is this informative about which prepa-
ration was performed on the other? The Preparation Uninformativeness Condition
is the condition that it is not. This condition is satisfied whenever the PIP is. It is
also satisfied if you regard the effects of the preparations to be ontically distinct: in
such a case, given the ontic state of the joint system, you are already certain about
which preparations have been performed, and being told about the preparation on
one system will not shift your credences.
458 W. C. Myrvold

The PUC is implied by the PIP, but it is strictly weaker. Even if the CPA is
assumed, it is possible to construct models for the PBR setup, outlined in the
previous section, in which the PUC is satisfied but the PIP is not. See Myrvold
(2018) for one such construction.
On the assumption of the Preparation Uninformativeness Condition—which,
in the current context, is a condition on the agent’s credences, independent of
the structure of the state space of the theory considered—one can get a ψ-
ontology proof for a class of nonorthogonal states. Let {φ i , ψ i } (i = 1, 2) be
preparation-acts such that Alice’s credences about the results of future experi- √
ments can be represented by quantum states {|φ i , |ψ i }. If |φ i |ψ i | ≤ 1/ 2,
then {|φ 1 |φ 2 , |φ 1 |ψ 2 , |ψ 1 |φ 2 , |ψ 1 |ψ 2 } is an antidistinguishable set. From
these assumptions, together with a mild side-assumption called the principle of
extendibility,3 it follows that, for any theory T such that Alice’s credences satisfy
the PUC for each of the preparation-acts {φ 1 ⊗ φ 2 , φ 1 ⊗ ψ 2 , ψ 1 ⊗ φ 2 , ψ 1 ⊗ ψ 2 },
Alice takes the effects of the preparation acts {φ i , ψ i } to be ontically distinct.
A key feature of quantum mechanics used in these proofs is √the fact
that, if {|φ i , |ψ i } are any state vectors with |φ i |ψ i | ≤ 1/ 2, then
{|φ 1 |φ 2 ,|φ 1 |ψ 2 ,|ψ 1 |φ 2 ,|ψ 1 |ψ 2 } is an antidistinguishable set. What is
needed for the argument is that there exist some preparations and an operation
such that Alice’s preparation-response conditional credences match the Born-rule
probabilities yielded by these states for the outcomes of the experiment envisaged.
This is not a necessity or probabilistic coherence. What we will assume is that Alice
accepts quantum mechanics, in the following sense: for any quantum state ρ of a
system, and any complete set of mutually orthogonal projections {P1 , . . . , Pn } on
the Hilbert space of the system, there is some combination of preparation-act and
readout-operation such that Alice’s credences in the possible results {r1 , . . . , rn }
match (or at least closely approximate) the Born-rule probabilities Tr(ρ Pk ).4
The upshot of these theorems is that, for any theory T for which the assumptions
(which may include both assumptions about the theory and about Alice’s credences
conditional on the supposition of the theory), Alice’s credences, conditional on T ,
should take the effects of the preparation-acts considered to be ontically distinct.
Thus, her credence in the ontic view of quantum states should be at least as high
as her total credence in the class of all such theories. The only way that she can
categorically deny that preparation-acts corresponding to distinct quantum states
yield ontically distinct states of reality is to attach credence zero to the class of all
such theories.

3 This says that any system composed of N subsystems of the same type can be regarded as a part
of a larger system consisting of a greater number of systems of the same type.
4 Of course, we could reasonably expect something stronger, that this holds, not just for projections,

but also for any positive-operator valued measure (POVM). But the weaker condition is all that we
need.
20 Quantum Probabilities 459

20.5 The QBist Response

Advocates of QBism will not accept the conclusion that a subjectivist about quantum
probabilities should be a realist about quantum states. Yet the argument so far has
not violated any of the core tenets of QBism. To avoid the conclusion, QBists must
add to its main tenets a further prohibition, logically independent of the explicitly
stated core tenets.
In section 2, we first argued for what, in my opinion, ought to be an uncontentious
claim: that an agent who regards a pair of preparation-acts to be distinguishable
ought to regard the effects of these acts to be ontically distinct. This has the
consequence that, if an agent associates orthogonal quantum states with a pair of
preparation-acts, she ought to take their effects to be ontically distinct. The extension
of this conclusion to nonorthogonal quantum states requires additional assumptions,
which might be questioned. But the QBist will want to resist the very first step,
having to do with orthogonal quantum states.
Nonetheless, the reasoning involved does not violate what proponents of QBism
have identified as its fundamental tenets. These are given in different versions in
different venues, but this is true for any of these versions.
Fuchs, Schack, and Mermin (2014) present the following as the “three funda-
mental precepts” of QBism.
1. A measurement outcome does not preexist the measurement. An outcome is
created for the agent who takes the measurement action only when it enters the
experience of that agent. The outcome of the measurement is that experience.
Experiences do not exist prior to being experienced.
2. An agent’s assignment of probability 1 to an event expresses that agent’s personal
belief that the event is certain to happen. It does not imply the existence of an
objective mechanism that brings about the event. Even probability-1 judgments
are judgments. They are judgments in which the judging agent is supremely
confident.
3. Parameters that do not appear in the quantum theory and correspond to nothing
in the experience of any potential agent can play no role in the interpretation of
quantum mechanics.
Regarding Precept 1: We have not assumed that the result of a readout-operation
(we have refrained from using the misleading term “measurement”) preexists
the operation, or that experiences exist prior to being experienced. Nor have we
presupposed that the result of such an operation merely indicates a preexisting
element of reality.
Regarding Precept 2: What we have concluded is that, if an agent takes two
preparation-acts to be distinguishable with certainty, the agent should take the
effects of those acts to be ontically distinct. The agent may be mistaken about this,
of course. For example: I may be convinced that I will be able to learn tomorrow
about my choices of key-strokes today, because I regard different choices of what
to write today as readout-distinguishable tomorrow. Accompanying this belief is a
460 W. C. Myrvold

belief that distinct choices of keystrokes today, if I save the file and don’t erase it,
have the effect of putting my laptop’s internal mechanism into physically distinct
states, which will have observationally distinguishable effects tomorrow. I could be
wrong about this, of course. It might be that my laptop is malfunctioning and that
there are, in fact, no lasting effects of my choices of keystrokes today. But—and
this is the crucial point—if I were to revise my judgment about the ontic effects of
my keystrokes, I would also revise my assessment of the readout-distinguishability
of my choices tomorrow. What I would not do is simultaneously maintain that my
choices today have no effect on the internal state of the laptop and that the result of
the readout operation tomorrow will be informative about those choices.
Regarding Precept 3: The conclusion that an agent who takes two preparation-
acts to be distinguishable should regard the effects of these acts to be ontically
distinct is quite general, and is not specific to quantum mechanics. It invokes
the sort of considerations that can be brought to bear on any sort of theorizing
about the world. It should also be emphasized that the mere introduction of some
sort of physical state-space Ω does not automatically consign us to the realm
of hidden-variables interpretations of quantum mechanics. The structure of the
physical state-space is left completely open, and our conclusion that one should
take there to be distinctions in physical reality corresponding to distinctions between
pure quantum state-preparations does not, by itself, commit us to anything else.
Fuchs (2017) presents the “three tenets” of QBism as
1. The Born rule is a normative statement.
2. All probabilities, including all quantum probabilities, are so subjective they never
tell nature what to do. This includes probability-1 assignments. Quantum states
thus have no “ontic hold” on the world.
3. Quantum measurement outcomes just are personal experiences for the agent
gambling on them.
Regarding Tenet 1: Neither the Born rule, nor anything at all quantum-mechanical,
was invoked in the argument for the conclusion that an agent should take the
effects of preparation-acts that she regards as readout-distinguishable to be ontically
distinct. A special case of this is a pair of preparations that are such that her
credences can be represented by a orthogonal quantum states. In the extension
to non-orthogonal quantum states, Born-rule probabilities were invoked only in
consideration of what sorts of credences an agent who accepts quantum mechanics
should have.
Regarding Tenet 2: Again, we have not assumed anything at all about what nature
does. Our conclusion has only to do with the meshing of an agent’s act-outcome
conditional credences with her credences about the states of the physical system
that mediates between the acts and the outcomes. As we have already mentioned
in connection with the malfunctioning laptop, that system is under no obligation to
conform to her credences, and, she may, indeed, find her expectations disappointed.
If the last sentence in the statement of Precept 2 is supposed to indicate a denial
of an ontic construal of quantum states, then, despite the “thus,” it does not follow
from the first. As we have seen, one can begin with the assumption that quantum
20 Quantum Probabilities 461

probabilities are subjective and end with a conclusion that an agent whose credences
conform to quantum mechanics ought to take quantum states to be ontic.
Regarding Tenet 3: Nothing in the argument depends on whether the results {rk }
of the operation are taken to be things like states of the laptop screen, or states of
the agent’s nervous system, or her conscious experience upon looking at the screen.
Given that we have not, in our arguments, violated the fundamental precepts or
tenets of QBism, why would, and how could, a QBist resist the conclusion? The only
option open is to reject the claim that there is a causal relation between my choices
of preparation-acts today and my experiences of the results of readout operations
tomorrow that is mediated by the effects of my actions on the state of the physical
system on which I act. There are two ways to do this. One is to maintain that there is
a direct causal link between my present actions and future experiences, unmediated
by any effect of my actions on the physical system on which I act. The other is to
reject all talk of causal links between my present actions and future experiences.
There are passages in some writings by QBists that suggest the latter.
There is a sense in which this unhinging of the Born Rule from being a “law of nature” in
the usual conception . . . i.e., treating it as a normative statement, rather than a descriptive
one — makes the QBist notion of quantum indeterminism a far more radical variety than
anything proposed in the quantum debate before. It says in the end that nature does what
it wants, without a mechanism underneath, and without any “hidden hand” (Fine 1989) of
the likes of Richard von Mises’s Kollektiv or Karl Popper’s propensities or David Lewis’s
objective chances, or indeed any conception that would diminish the autonomy of nature’s
events. Nature and its parts do what they want, and we as free-willed agents do what we can
get away with. Quantum theory, on this account, is our best means yet for hitching a ride
with the universe’s pervasive creativity and doing what we can to contribute to it (Fuchs
2017, 272).

The suggestion is that the physical world obeys neither deterministic nor stochastic
laws.
What sort of physical laws, if any, nature conforms to is not something that can
be known a priori. If someone proposes a physical theory according to which one
should expect to observe certain statistical regularities, that theory can be subjected
to empirical test according to the usual Bayesian methods. This will never produce
certainty that the theory is correct, but can result in high credence that the theory,
or something like it, is at least approximately correct within a certain domain of
application. In particular, if someone were to propose a theory according to which
a given sort of preparation-act (say, passing a particle of a certain type through a
Stern-Gerlach apparatus oriented vertically, and selecting the + beam) bestowed
determinate chances on certain sorts of subsequent experiments, we could test that
assertion by repeating the preparation-experiment sequence, and looking to see
whether the relative frequencies converged to a stable value, in accordance with the
weak law of large numbers. If we routinely find the theoretical expectation fulfilled,
this should boost credence in the theory; if the outcomes of the experiment are not
at all like what the theory leads us to expect, this counts as evidence against the
theory. One could imagine that persistent failure to find any sort of regularity at all
might boost credence in the proposition that there are no regularities to be found—
462 W. C. Myrvold

if the agent could survive that long! (It may be a requirement for the existence of
agents that could have credences that there be some modicum of predictability in
the world.)
Despite passages like the one quoted above, I am skeptical that any QBist actually
rejects the claim that there is a causal link between their present actions and their
future experiences, mediated by an effect of their actions on the physical systems
with which they interact. I find it difficult to imagine how someone whose beliefs
were like that would act. I am certain, however, that someone with beliefs like that
would not judge a laptop that was advertised as having more memory—that is, more
readout-distinguishable internal states that the user could choose between via choice
of actions performed on the machine—to be worth more money than one with less
memory. Indeed, it is hard to see why an agent like that could prefer any laptop at
all to a kumquat as an implement for taking notes.5
Fuchs does not explain what, if anything, he takes to be evidence for the radical
indeterminism of which he speaks. Such evidence cannot have come from the
results of quantum experiments, which, on the contrary, seem to indicate law-like
connections, best expressed in terms of probabilistic laws, between preparation-acts
and experimental outcomes. At any rate, the evidence against the claim that the
world is so completely devoid of law-like regularities that theorizing about it is
pointless seems to be overwhelming.
There is, it seems, one possible response for the QBist. This would be to say
that he accepts the thesis of radical indeterminism, of a world subject neither to
deterministic nor stochastic laws, not on the basis of evidence, but as a leap of faith,
or an expression of temperament (see Fuchs 2017, 251–253). Though he believes in
no causal connection between his actions on the world and future experiences, as a
psychological weakness he is unable to set his credences about the future at will, and
can only affect his credences about the future by undergoing a ritual manipulation
of the physical objects around him. He finds himself in the position of someone who
does not believe that a horseshoe hung on the door will bring good luck, but hangs it
anyway because he finds that doing so makes him more optimistic about the future.
If that is the view—namely, that a QBist can offer no reason whatsoever to accept
the radical renunciation of physical theorizing about the world that acceptance of
the view entails—then we are in full agreement on that point.

5 Itis worth re-emphasizing: I am not claiming that anything about the laptop or the kumquat
follows from assertions about the agent’s credences about these systems. The agent may prefer
the laptop to the kumquat as a note-taking device because of a false belief that the laptop is more
suitable to play that role in a causal chain between the agent’s actions and future experiences. The
point is merely that the agent’s belief, true or false, is a belief about the physical workings of the
laptop.
20 Quantum Probabilities 463

20.6 An Argument from Locality?

In addition to the tenets and precepts listed above, QBism also involves a severe
restriction on the application of quantum mechanics. An agent is required to only
apply quantum probabilities to her own future experiences, and not to events
(including the experiences of other agents) that she herself will not experience.
The severity of this restriction of scope is not always sufficiently emphasized.
Though some of the ideas adopted by QBists are inspired by quantum information
theory, the restriction of credences to one’s own experiences eliminates vast swaths
of quantum information theory, including everything to do with communication and
cryptography. The point of a communication is to influence the credences of another
agent, and, in designing a communication protocol, one must consider probabilities
of changes to the recipient’s belief-state.6
There may seem to be an advantage to this restriction: one thereby avoids
quantum nonlocality. If an agent uses quantum mechanics to assign probabilities
to events along her own worldline, and never to events at a distance from herself, all
the events to which she assigns probabilities are timelike-related. There can be no
question of action-at-a-distance because there are no distances between the events
she assigns credences to.
This advantage obtains only if one refrains from doing any theorizing, quantum
or otherwise, about events at a distance. Any theory, quantum or otherwise, that
attempts to account for outcomes of tests of Bell inequalities—if we mean by
‘outcomes’ what is usually meant, namely, detector-registrations that occur at
spacelike separation from each other—will have to violate at least one of any set
of conditions sufficient to yield Bell inequalities.
The seeming advantage is illusory. If one refrains from giving an account of
goings-on in the world that occur at a distance from each other in order to avoid
giving an account of nonlocal goings-on, and if it is held that such a renunciation
is necessary to avoid some sort of nonlocality that is thought to be objectionable,
this is tantamount to saying that if one were to give an account of goings-on in
the world that occur at a distance from each other, such an account would involve
the objectionable sort of nonlocality. If this differs from acknowledging that the
supposedly objectionable nonlocality is a feature of the world, I don’t know how. It
is as if one declined a request to evaluate a co-worker by saying that one is declining

6 Itmight be claimed, in response, that Alice, in sending a signal to Bob, is concerned about the
impact on Bob only insofar as it will later impinge on her. This, as everyone knows, is false. Some
of the intended results of a communication are ones that the sender will never be aware of. As a
particularly vivid example, consider a spy whose cover has been blown, who sends out one last
encrypted message before taking a cyanide pill to avoid capture. A QBist must regard this as a
misuse of quantum mechanics if the encryption scheme is a quantum scheme. It is not! Nor is it
a misuse of quantum mechanics for an engineer designing a nuclear waste storage facility to use
quantum mechanics to compute half-lives with the aim of constructing a facility that will be safe
for generations to come.
464 W. C. Myrvold

the request in order to avoid saying that the co-worker is incompetent. In giving such
a response, one has not refrained from conveying one’s judgment of the co-worker’s
competence.

20.7 Conclusion

Subjectivitism about probabilities is not, by itself, an escape from ψ-ontology

theorems. The theorems go through with all probabilities interpreted subjectively,
with a conclusion that an agent whose credences satisfy the conditions of the proof
should believe that preparation procedures that she associates with distinct pure
quantum states leave the systems on which those procedures are performed in
distinct ontic states. If this conclusion is to be avoided, some other stricture must
be placed on the agent’s credences.
The chief condition that underlies the proof is the assumption of the possibility of
performing a preparation procedure that effectively screens off correlations between
the system on which it is performed and the rest of the world, rendering a further
specification of its physical state uninformative about the states of other systems.
This sort of assumption lies deep at the heart of virtually all experimental science.
None of the empirical evidence that motivates a shift from classical to quantum
theory provides grounds for rejecting assumptions of this sort. If we ever come to
a point at which we have reasons to doubt this sort of assumption, this will not
come about as a result of experiments that presuppose it. And if we were presented
with a reason to doubt this sort of assumption, it is hard to see how this doubt
could be sufficiently contained so as not to undermine all of experimental science.
Fortunately, we are not in such a position.

References

Caves, C. M., Fuchs, C. A., & Schack, R. (2002). Quantum probabilities as Bayesian probabilities.
Physical Review A, 65, 022305.
Caves, C. M., Fuchs, C. A., & Schack, R. (2007). Subjective probability and quantum certainty.
Studies in History and Philosophy of Modern Physics, 38, 255–274.
Clifton, R., & Halvorson, H. (2000). Bipartite-mixed-states of infinite-dimensional systems are
generically nonseparable. Physical Review A, 61, 012108.
Clifton, R., & Halvorson, H. (2001). Entanglement and open systems in algebraic quantum field
theory. Studies in History and Philosophy of Modern Physics, 32, 1–13.
Fine, A. (1989). Do correlations need to be explained? In J.T. Cushing and E. McMullin (Eds.),
Philosophical Consequences of Quantum Theory: Reflections on Bell’s Theorem. Notre Dame,
IN: University of Notre Dame Press, Notre Dame, 175–194.
Fuchs, C. A. (2017). Notwithstanding Bohr, the reasons for QBism. Mind & Matter, 15, 245–300.
Fuchs, C. A., & Schack, R. (2013). Quantum Bayesian coherence. Reviews of Modern Physics, 85,
1693–1715.
20 Quantum Probabilities 465

Fuchs, C. A., & Stacey, B. (2019). QBism: Quantum theory as a hero’s handbook. In E. M.
Rasel, W. P. Schleich, & S. Wölk (Eds.) Foundations of quantum theory: Proceedings of the
international school of physics “Enrico Fermi” course 197, pp. 133–202. Amsterdam: IOS
Press.
Fuchs, C. A., Mermin, N. D., & Schack, R. (2014). An introduction to QBism with an application
to the locality of quantum mechanics. American Journal of Physics, 82, 749–754.
Harrigan, N., & Spekkens, R. W. (2010). Einstein, Incompleteness, and the Epistemic View of
Quantum States. Foundations of Physics, 40, 125–157.
Leifer, M. S. (2014). Is the quantum state real? An extended review of ψ-ontology theorems.
Quanta, 3, 67–155.
Myrvold, W. C. (2018). ψ-ontology result without the Cartesian product assumption. Physical
Review A, 97, 052109.
Myrvold, W. C. (2020). On the status of quantum state realism. In S. French & J. Saatsi (Eds.)
Scientific realism and the quantum. Oxford: Oxford University Press, 229–251.
Pitowsky, I. (2003). Betting on the outcomes of measurements: A Bayesian theory of quantum
probability. Studies in History and Philosophy of Modern Physics, 34, 395–414.
Pusey, M. A., Barrett, J., & Rudolph, T. (2012). On the reality of the quantum state. Nature Physics,
8, 475–478.
Chapter 21
The Relativistic Einstein-Podolsky-Rosen
Argument

Michael Redhead

Abstract We present the possibility of a relativistic formulation of the Einstein-

Podolsky-Rosen argument. We pay particular attention to the need for a refor-
mulation of the so-called reality criterion. We introduce such a reformulation for
the reality criterion due to Ghirardi and Grassi and show how it applies to the
nonrelativistic EPR argument. We elaborate on Ghirardi and Grassi’s proof and
explain why it cannot be circumvented. Also we draw a new conclusion by assuming
locality. Finally, we review and summarise our own views. This is a continuation
of the paper by myself and Patrick La Rivière whose defects are exposed, and
corrected, in the present work. In the Appendix locality principles are conveniently
spelled out.

Keywords New critique of Ghirardi and Grassi’s paper assuming locality · New
refutation of Redhead and Rivière’s paper · Nonlocality issues revisited

21.1 Preliminaries

At first glance, the realist interpretations of quantum mechanics such as Bohm’s

(see Bohm and Hiley 1993) offer many advantages over standard interpretations
of the theory. In particular, they give a clear, intuitive picture of many poten-
tially paradoxical, physical situations, such as the two-slit experiment and the
phenomenon of barrier penetration. At the same time, their chief drawback—a form
of nonlocality, as we have claimed, to conflict with the constraints of relativity
theory – is apparently shared by the standard ‘anti-realist’ interpretations that reject
hidden variables and assume completeness, as was demonstrated by the original
Einstein-Podolsky-Rosen (EPR) argument (see Einstein et al. 1935).

M. Redhead ()
Centre for Philosophy of Natural and Social Science, London School of Economics and Political
Science, London, UK
e-mail: mlr1000@cam.ac.uk

© Springer Nature Switzerland AG 2020 467

M. Hemmo, O. Shenker (eds.), Quantum, Probability, Logic,
Jerusalem Studies in Philosophy and History of Science,
https://doi.org/10.1007/978-3-030-34316-3_21
468 M. Redhead

While the Bell argument that establishes nonlocality for realistic interpretations
such as Bohm’s has been formulated in an experimental relativistic context (see
Bohm and Hiley, Chap. 12, 1993) there is no well-established relativistic formula-
tion of the EPR argument. In the absence of such a formulation, it seems hasty to
conclude that the tension between the standard interpretations and relativity theory
is just as great as that between detailed Bohmian interpretations and relativity.
Clearly, if a relativistic formulation of EPR could be given that did not entail
nonlocality, antirealist interpretations would have an advantage over the Bohmian
interpretation. (To be sure anti-realist interpretations that do not require nonlocality
include Healey’s pragmatism (see Healey 2017), Qbism (see Caves et al. 2002) and
epistemicism (see Williamson 1994).
Let us now investigate the possibility of a relativistic formulation of the
EPR argument. First, the standard nonrelativistic version of the EPR argument
is reviewed and the problematics of translating it into a relativistic context are
considered, paying particular attention to the need for a reformulation of the so-
called reality criterion. Then, we introduce one such reformulated reality criterion,
due to GianCarlo Ghirardi and Renata Grassi (see Ghirardi and Grassi 1994) and
show how it is applied to the nonrelativistic EPR argument. Next, we discuss the
application of the new reality criterion in a relativistic context.
A relativistic version of the EPR argument must differ from the nonrelativistic
version in two principal ways. First, the particle states must be described by a
relativistic wave-function. The details don’t concern us here; we need only require
that the wave-function preserve the maximal, mirror-image correlations of the
nonrelativistic singlet state. And indeed, the existence of maximal correlations in
the vacuum state of relativistic algebraic quantum field theory has recently been
demonstrated (see Redhead 1995). Second, the argument must not depend on the
existence of absolute time ordering between the measurement events on the left and
right wings of the system, for in the relativistic argument these may be spacelike
separated. As it turns out, the nonrelativistic version of the argument does invoke
absolute time ordering. To see how to get around this problem, we must briefly
review the standard formulation of the incompleteness argument.
For EPR, a necessary condition for the completeness of a theory is that every
element of physical reality must have a counterpart in the theory. To demonstrate
that quantum mechanics is incomplete, EPR need simply point to an element of
physical reality that does not have a counterpart in the theory. In this vein, they
consider measurements on a pair of scattered particles with correlated position
and momentum, but the formulation of the argument in Bohm (see Bohm 1951)
in terms of a pair of oppositely moving, singlet-state, spin-1/2 decay products of
a spin-0 particle, is conceptually simpler. In this case, the formalism of quantum
mechanics demands a strict correlation between the spin components of the two
spatially separated particles, such that a measurement of, say, the z-component of
spin of one particle allows one to predict with certainty the outcome of the same
measurement on the distant particle . This ability to predict with certainty, or at least
probability one, the outcome of a measurement is precisely the EPR criterion for the
existence of an element of reality at the as-yet-unmeasured particle. By invoking one
21 The Relativistic Einstein-Podolsky-Rosen Argument 469

final assumption, a locality assumption stating that elements of reality pertaining

to one system cannot be affected by measurements performed ‘at a distance’ on
another system, EPR can establish that the element of reality at the unmeasured
particle must have existed even before the measurement was performed at the distant
particle . But the quantum formalism describes the particles at this point with the
singlet state, and thus has no counterpart for the element of reality at the unmeasured
particle. It follows that the quantum description was incomplete. Schematically,

Quantum Formalism ∧ Locality →∼ Completeness. (21.1)

Alternatively, if one assumes completeness, the argument may be rearranged as

a proof of nonlocality:

Quantum Formalism ∧ Completeness →∼ Locality. (21.2)

The problematic assumption of absolute time ordering entered the argument in

the reality criterion, which turns on the possibility of predicting with certainty the
outcome of a measurement along one wing subsequent to having obtained the result
of a measurement along the other. Of course, for spacelike separated events, notions
like precedence and subsequence are reference-frame dependent, not absolute. So to
translate the EPR argument to a relativistic context requires a modified criterion for
the attribution of elements of reality that is not contingent on the time ordering of
the measurement events. Ghirardi and Grassi have undertaken to formulate just such
a criterion. For the sake of clarity, we shall first describe how this criterion applies
to the nonrelativistic version of the argument.
Ghirardi and Grassi’s criterion rests on the truth of certain classes of counter-
factual statements-statements of the form ‘if ϕ were true, then ψ would be true’
where the antecedent ϕ is in general known to be false. In particular, they wish
to link the attribution at a time t of the property corresponding to [observable a
having value a] to the truth of the counterfactual assertion: if measurement of a
were performed at time t, then the outcome would be a (they use David Lewis’s
analysis of counterfactuals here. See Lewis 1986).
With this criterion in hand, Ghirardi and Grassi can now run the nonrelativistic
EPR argument essentially as before. They assume a measurement of property a
is performed on the right-hand particle at time tR , yielding a specific result a. To
ascertain whether an element of reality corresponding to property a = a exists at
the left-hand particle, they must assess the truth of the counterfactual assertion: ‘if
I were to perform a measurement of property a at the left-hand particle at time tL , I
would obtain the result a .’ In the nonrelativistic case, the truth of this counterfactual
assertion follows naturally from the presence of absolute time ordering. For if
tR < tL , then the outcome of the right-hand measurement can be assumed to be
the same in all of the ‘accessible’ (most similar) worlds used to evaluate the
counterfactual, because it is strictly in the past of the counterfactual’s antecedent.
The strict correlation laws of quantum mechanics, also assumed to hold in all
accessible worlds, then demand that the result of a measurement on the left wing also
470 M. Redhead

be fixed in all possible worlds (specifically, the laws require that a = −a). Thus the
counterfactual is true, and an element of reality can be said to exist at the left-hand
particle. From here, the argument unrolls in the usual way, and by supplementing
this reality criterion with a locality assumption (they call it G-Loc, after Galileo),
Ghirardi and Grassi can deduce that quantum mechanics is incomplete. Once again,
we can represent their argument schematically by

Quantum Formalism ∧ G-Loc →∼ Completeness (21.3)

or

Quantum Formalism ∧ Completeness →∼ G-Loc. (21.4)

21.2 The Crux of the Argument

‘A system cannot be affected by actions on a system from which it is isolated. In

particular, elements of physical reality of a system cannot be influenced by actions
on systems from which it is isolated.’ An examination of the structure of Ghirardi
and Grassi’s argument reveals that they make use not of the general principle stated
but of a special case of this general principle, namely that elements of reality cannot
be brought into existence ‘at a distance.’ It is this special case of G-Loc, call it ER-
Loc (for elements of reality) that enters toward the end of the argument to establish
that the measurement at the right wing could not have created an element of reality
at the left wing and thus it must have existed prior to the measurement at the right
wing, when the quantum formalism said the particles were in the singlet state. Thus
they conclude that quantum mechanics is incomplete. All is well so far, but when
one turns the argument around, assuming completeness and dispensing with locality,
one must ask, can one be more precise as to which locality principle should be given
up: the principle they label G-Loc, or the special case ER-Loc? Indeed it is the latter,
for only it entered into the argument. As it turns out the distinction between G-Loc
and ER-Loc does not affect their conclusions in the nonrelativistic case, because the
conclusion they choose to highlight- the creation of elements of reality at a distance-
is precisely one that does follow from dispensing only with ER-Loc.
In the relativistic case they want to claim peaceful coexistence between nonlocal-
ity and special relativity. However, greater care must be taken with the statement of
the locality principle, this time called L-Loc (after Lorentz by Ghirardi and Grassi),
because a locality principle must enter at the very beginning of the argument as
well as in the usual way at the end. The argument begins in the same way as
in the nonrelativistic case, with the occurrence of a measurement on the right-
hand side, but now the absence of absolute time ordering means the result of this
measurement can no longer tacitly be assumed to be the same in all the accessible
worlds used to evaluate the element-of-reality counterfactual at the lefthand side.
Locality must be invoked to establish the independence of the outcome of the
21 The Relativistic Einstein-Podolsky-Rosen Argument 471

righthand measurement from the occurrence of the measurement at the left. This
done, Ghirardi and Grassi then demonstrate the existence of an element of reality at
the left-hand side following the same reasoning as above. From here, the argument
unrolls once again in the usual way and locality makes a second appearance in its
familiar place at the end of the argument. In this way, Ghirardi and Grassi can again
prove that standard quantum mechanics plus ‘locality’ implies incompleteness.
But there are two quite distinct cases of L-Loc that are actually being employed,
one used in getting the argument started and the other appearing in the conclusion.
Ghirardi and Grassi define L-Loc as the following: ‘An event cannot be influenced
by events in spacelike separated regions. In particular, the outcome obtained in
a measurement cannot be influenced by measurements performed in spacelike
separated regions; and analogously, possessed elements of physical reality referring
to a system cannot be changed by actions taking place in spacelike separated
regions.’ As in the nonrelativistic case, it is not the general principle but rather the
two special cases, call them OM-Loc (for outcome of measurement) and ER-Loc
(again for element of reality), that are doing the logical work in their argument. OM-
Loc affirms that the outcome of a measurement cannot be influenced by performing
another measurement at a spacelike separation, while ER-Loc affirms that elements
of reality cannot be created by performing a measurement at spacelike separation.
Ghirardi and Grassi invoke OMLoc at the beginning of the argument while applying
the counterfactual reality criterion, as discussed above, and they invoke ER-Loc
at the end of the argument, as they did in the nonrelativistic case. So if we write
L-Loc = OM-Loc ˆ ER-Loc, then, schematically, their argument looks like this:

Quantum Formalism ∧ OM-Loc ∧ ER-Loc →

(21.5)
∼ Completeness

or

Quantum Formalism ∧ Completeness →∼ OM-

(21.6)
Loc v ∼ ER-Loc.

Ghirardi and Grassi now argue, in effect, as follows. Assuming OM-Loc we can
again demonstrate from Completeness a violation of ER-Loc, i.e. Einstein’s spooky
action-ata-distance creating elements of reality at a distance. But if we don’t assume
OM-Loc, then we cannot deduce a violation of ER-Loc. All this is quite correct, but
the price we have to pay for not being able to demonstrate a violation of ER-Loc is
precisely that we have to accept a violation of OM-Loc!
In other words, the relativistic formulation of the EPR argument does not
help with the thesis of peaceful coexistence between quantum mechanics and
special relativity, unless one argues that violating ER-Loc is more serious than
violating OM-Loc from a relativistic point of view. This is hard to maintain since
violating OM-Loc involves a case-by-case version of what Shimony (see Shimony
1984) refers to as violating parameter independence, i.e. the independence of the
probability of the outcome on one wing of the EPR experiment with respect to the
472 M. Redhead

parameters controlling the type of experiment being performed on the other wing.
By analogy, violating ER-Loc is also a form of parameter dependence.
So to return again, Ghirardi and Grassi’s claim of peaceful coexistence is already
justified, with a novel idea of how it works out, a violation of OM-Loc.

21.3 The Thesis of Myself and La Rivière’s Paper

However they say they need to identify an additional assumption omitted from
(21.5) and (21.6), which, if challenged, could undermine the inference. Recall that
to run the argument in either the nonrelativistic or relativistic case, Ghirardi and
Grassi must establish that the outcome of, say, the right-hand measurement is the
same in all accessible worlds. With this established, the correlation laws of quantum
mechanics imply that the outcome of the left-hand measurement is the same in all
accessible worlds, and hence establish the truth of the counterfactual assertion about
the left-hand measurement result that permits the attribution of an element of reality
to the left hand particle. In the nonrelativistic case, the constancy of the right-hand
result is a natural consequence of the absolute time ordering as discussed above; in
the relativistic case. A premise akin to one that Michael Redhead labels the Principle
of Local Counterfactual Definiteness (PLCD) (see Redhead, p. 92, 1989) is needed
to do this sort of work.
In the present case, PLCD may be taken to assert that the result of an experiment
that could be performed on a microscopic system has a definite value that does
not depend on the occurrence of a measurement at a distant apparatus. Ghirardi
and Grassi implicitly assume that PLCD is licensed by their locality principle, for
they invoke only OM-Loc to establish the constancy of the right-hand outcome in
all accessible worlds. But we argue that PLCD does not follow directly from any
typical locality principle, certainly not from one like OM-Loc, which asserts that
the outcome obtained in a measurement cannot be influenced by measurements in
spacelike separated regions. The reason is quite simple: while invoking locality may
prevent measurements on the left hand particle from influencing the result at the
right and from breaking the constancy of the accessible worlds as far as the right-
hand result is concerned, it does not prevent indeterminism from wreaking that sort
of havoc. Intuitively, we can imagine that we run the world over again, this time
performing the measurement on the left-hand particle. If we consider this left-hand
measurement schematically as a point event with a backward light cone identical to
that in the actual world, we are concerned with what will happen in the complement
of the forward and backward light cones. Under indeterminism the events in this
complement (the absolute elsewhere) simply cannot be assumed to remain the same.
Thus it is maintained that Ghirardi and Grassi need both OM-Loc and an
assumption of determinism to get their argument off the ground.
21 The Relativistic Einstein-Podolsky-Rosen Argument 473

We need, schematically, to replace (21.6) by

Quantum Formalism ∧ Completeness ∧

(21.7)
Determinism →∼ OM-Loc v ∼ ER-Loc.

Thus,
Determinism ➔ OM-Loc (➔ PLCD)

and hence by the disjunctive syllogism, we can infer ~ER-Loc,

So text of myself and La Rivière (see Redhead and La Rivière 1997) ends here.

21.4 But there Is a Catch . . .

But there is a catch. We note first that Determinism ➔ Bell Inequality (assuming
randomness assumption is assumed to be true, for which independent engagement
can be given). But this contradicts the conclusion that the Bell Inequality is false.
So we cannot infer ~ER-Loc.
But, according to Lewis (private communication), OM-Loc ➔ PLCD and hence
allows the Bell Inequality, independently of Determinism. The argument here
appears to be the following. For possible worlds we allow so-called quasi-miracles
to happen, but also they might not happen. To prove the counterfactual we need to
close down worlds where miracles happen, leaving the counterfactual to succeed.
We disagree with this claim since it involves discussion of possible worlds,
whereas in our world the counterfactual comes out as false.
Put succinctly, the argument is the following. A counterfactual supposition about
an event can, and indeed should, be taken to preserve as much as possible of the
actual facts, compatible with the truth of the supposition. Since in a relativistic
setting, a counterfactual supposition about an event is compatible with preserving
the the actual facts that are separated from it. It seems to us, under indeterminism,
one’s treatment of a counterfactual supposition should ‘respect’ the indeterminism
by allowing the ways that history could play out, compatibly with the counterfactual
supposition to be considered.
But if we follow Lewis we again allow the Bell Inequality so on both counts the
attempt to rescue the justification of ~ER-Loc fails.
We make a remark on Ghirardi and Grassi’s proof on peaceful existence. For
Ghirardi and Grassi there are two sorts of argument: (1) EPR argument assuming
locality; (2) argument showing, correctly, how nonlocality works in violating Bell.
I confused the two sorts of arguments, starting with (1) and trying to prove peaceful
existence according to (2)! I learned from my mistakes as Popper would have said!
But notice how we have shown, using (1), that peaceful coexistence actually
works out.
474 M. Redhead

21.5 Conclusion

So what conclusion concerning nonlocality can we draw? Taking first the realist
option, in which all observables have sharp values at all times, if we assume
determinism or restrict the discussion to cases of strict correlation or anticorrelation
where we can derive determinism on plausible assumptions, there seems very little
scope for avoiding the conclusion of nonlocality for any ‘realist’ reconstruction
of quantum mechanics. But experimentally we never observe absolutely strict
correlations, so it may be argued that discussions of nonlocality should not deal
with this ideal case, but should be based upon the experimentally realistic nonideal
case. We are not convinced by this argument, since idealization is an essential aspect
of any scientific theorizing. Be that as it may, the assumption of determinism for
the nonideal case where correlations are less than perfect, may well be suspect. In
such cases one is forced, as we have seen, to go to the stochastic hidden-variable
framework, and much of the recent discussion has focused on the significance of
violating the locality assumptions involved in this framework. Following Jon Jarrett
(Jarrett 1984) these break down into two classes.

1. Independence of the probability for a particular outcome of measurement on one particle,

conditionalized on the collective hidden variables, from the outcome of measurement
carried out on a remote particle. This is what Shimony calls parameter independence,
which is already referred to above.
2. Independence of the probability for a particular outcome of measurement on one particle
conditionalized on the collective hidden variables, from the type of measurement being
carried out on a remote particle. This is what Shimony calls outcome independence.
(Shimony 1984).

Violation of (1) has been generally held to mark probabilistic causal dependence
between the outcome of a local measurement and the ‘setting’ of a remote piece of
apparatus, in prima facie conflict with standard interpretations of special relativity
as prohibiting causal links between spacelike separated events. Violation of (2) has
led to more controversy. It seems to demonstrate that the precise specification of the
state of the source, or more generally of the particles before the measurements are
undertaken, cannot be regarded as a common cause of the measurement outcomes.
Indeed a direct causal connection exists between the events consisting of the
measurement results on distant particles obtaining in the way they actually do. We
have challenged this view (Redhead, pp.90–96, pp.98–107, 1989) by proposing to
interpret violation of (2), in the way proposed, but in terms of a noncausal direct
dependence between measurement outcomes that lacks the necessary ‘robustness’ in
respect of how the measurement results are brought about, to merit description as a
causal connection. This approach gives up the stochastic hidden-variable framework
in favour of trying to understand the correlations between distant events in terms of a
harmony-at-a distance, or, as Shimony describes it, a passionat-a distance (Shimony
1984).
This idea ties in with, but is really quite distinct from, the fact that violation
of (2) cannot be used to transmit signals between distant locations (Redhead,
21 The Relativistic Einstein-Podolsky-Rosen Argument 475

b p.106, 1989) since the necessity to recover quantum probability distributions by

averaging over the hidden variables, means that we have no way of controlling
the local marginal distributions for the results of measurement on one particle, by
varying the type of measurement performed on distant particles. The no-signalling
result is often cited to defuse the tension between nonlocality in quantum mechanics
and the constraints of special relativity, but, at a deeper level, the ‘nonrobustness’
argument may be preferable.
Turning to the anti-realist option, we have examined Ghirardi and Grassi’s
attempt to reformulate the EPR argument in a relativistic context and argued that
peaceful coexistence does arise.
There remains of course the question of how to interpret the violation of outcome
independence. Assuming completeness, outcome dependence famously follows. In
the EPR set-up this means that when measurements are performed at spacelike
separation on the two wings of the experiment, the results are mirror-image
correlated. As one potentiality gets actualized on the left, say, how does this happen
exactly in tandem with the opposite result on the right? Are we faced with a causal
effect, namely result-to-result causation, so that peaceful coexistence with relativity
is still challenged? Similar answers involving no-signalling or nonrobustness may
be provided as in the discussion above of the ‘realist’ option.
But it must be stressed that the mysterious harmony of the result-to-result corre-
lations remains arguably ‘spooky’ even if it does not involve causal dependence.
Another distinguishing feature of this harmony is its symmetrical character,
quite unlike the asymmetry that one would normally want to ascribe to a causal
connection. Shimony’s phrase ‘passion-at-a -distance’ seems exactly the right one
to capture what is going on, even if one concedes that the mystery of the EPR
correlations is not eliminated merely by introducing an apt nomenclature.
For the anti-realist the role of measurement is to actualize potentialities. If there
are no measurements, then there are no actualities! This observation is particularly
relevant to cosmological applications of quantum mechanics, where there is nothing
‘outside’ the universe to serve as a measuring device! So, from the cosmological
perspective, one can argue that the realist option is after all to be preferred. This
assumes the universe is finite in spacial extent. For an infinite universe the argument
fails. I owe this comment to my erstwhile student Tian Yu Cao.
Here one can argue that the nonseparability approach, as exemplified in blocking
the nonlocality proof by denying the 0-Loc principle, may in the end be the best way
of understanding the peculiar features of entangled states in quantum mechanics.
Many people find it difficult to see the distinction between violating 0-Loc
(nonseparability) and violating E-Loc (actionat-distance). A simple example due
to David Lewis (private communication) may help here. Consider someone with
what she calls a bilocal hand. There is in reality just one hand that is manifested
in two different places. If you shake hands with this curiously disabled person, her
other hand will move in synchrony with the one you are shaking, not because there
is an interaction between the two hands, but just because, in reality, there is only one
hand, bilocally located! According to Lewis, this is an example of nonseparability
as distinct from action-at-distance.
476 M. Redhead

Acknowledgements I acknowledge Jeremy Butterfield and Bryan Roberts for reading the
manuscript, and for helpful comments.

A.1 Appendix

A.1.1 Realist Interpretations

E-Loc: Local elements of reality cannot be affected by changes in a distant,

spacelike separated environment.
O-Loc: Local elements of reality can be specified independently of a holistic
context.

A.1.2 Anti-Realist Interpretations

G-Loc: Events cannot be affected by a distant measurement performed simultane-

ously.
L-Loc: Same as for G-Loc but with ‘simultaneously’ replaced by ‘at spacelike
separation’.
ER-Loc: Elements of reality cannot be created by distant measurements performed
simultaneously/at spacelike separation – a special case of G-Loc/L-Loc.
OM-Loc: The outcome of a measurement at one location cannot be affected by a
distant measurement performed at spacelike separation – another special case of
L-Loc.
PLCD: The outcome of a measurement that could be performed has a definite value
independent of whether or not another measurement takes place at a distant
spacelike separated location. OM-Loc implies PLCD under an assumption of
determinism, but not of indeterminism.
Parameter independence: The statistics of measurement results at one location is
independent of the parameters defining a measurement procedure at a distant
spacelike separated location.
Outcome independence: The statistics of measurement results at one location is
independent of the outcome of measurements performed at a distant spacelike
separated location.

References

Bohm, D. (1951). Quantum theory. Englewood Cliffs: Prentice-Hall.

Bohm, D., & Hiley, B. J. (1993). The undivided universe. Routledge: Taylor & Francis.
Caves, C. M., Fuchs, C. A., & Shack, R. (2002). Unknown quantum states: The de Finnetti
representation. Journal of Mathematical Physics, 43, 4537–4559.
21 The Relativistic Einstein-Podolsky-Rosen Argument 477

Einstein, A., Podolsky, B., & Rosen, R. (1935). Can quantum-mechanical description of physical
reality be considered complete? Physical Review, 47, 777–800.
Ghirardi, G., & Grassi, R. (1994). Outcome predictions and property attribution: The EPR
argument reconsidered. Studies in History and Philosophy of Science, 25, 397–423.
Healey, R. (2017). The quantum revolution in philosophy. Oxford: Oxford University Press.
Jarrett, J. (1984). On the physical significance of the locality conditions in the bell arguments.
Nous, 18, 569–589.
Lewis, D. (1986). Philosophy papers (Vol. 2). Oxford: Oxford University Press.
Redhead, M. L. G. (1989). Incompleteness, nonlocality and realism. Oxford: Clarendon Press.
Redhead, M. L. G. (1995). More ado about nothing. Foundations of Physics, 25, 123–137.
Redhead, M. L. G., & La Rivière, P. (1997). The EPR relativistic argument. In R. Cohen, M. Horne,
& J. Stachel (Eds.), Potentiality, entanglement and passion-at-a-distance. Quantum Mechanical
Studies for Abner Shimony (Vol. 2, pp. 207–215). Dordrecht: Kluwer.
Shimony, A. (1984). Controllable and uncontrollable non-locality. In S. Kamafuchi (Ed.), Pro-
ceedings of the international symposium: Foundations of quantum mechanics in the light of
new technology (pp. 225–230). Tokio: Physical Society of Japan.
Williamson, T. (1994). Vagueness. London: Routledge.
Chapter 22
What Price Statistical Independence?
How Einstein Missed the Photon

Simon Saunders

Abstract Einstein’s celebrated 1905 argument for light quanta was phrased in
terms of the concept of ‘statistical independence’– wrongly. It was in fact based
on the notion of local, non-interacting entities, and as such was neutral on the
question of statistical independence. A specific kind of statistical dependence had
already been introduced by Gibbs, in connection with the Gibbs paradox, but went
unrecognised by Einstein. A modification of Einstein’s argument favours the latter.
This kind of statistical dependence, as eventually attributed by Einstein to Bose,
although expressed, in quantum mechanics, by the (anti)symmetrisation of the
wave-function, does not require ‘genuine’ entanglement. This chapter demystifies
quantum indistinguishability; its novel features derive from the finiteness of state-
space in quantum mechanics not indistinguishability.1

Keywords Einstein · Boson · Photon · Gibbs paradox

22.1 Overview

According to the physicist and historian Abraham Pais, the photon was the first
particle to be predicted theoretically. It was also, he noted, met with scepticism on
an epic scale: ‘[i]ts assimilation came after a struggle more intense and prolonged
than for any other particle ever postulated’.2 Yet it was also, in a certain sense,
derived from empirical data – specifically, derived by Einstein from the Wien limit
of Planck’s black-body spectral distribution, taken as a phenomenological given

1 This chapter is based on Saunders (2018, 2020), but it’s focus is different from either.
2 Pais (1982, p. 361).

S. Saunders ()
Faculty of Philosophy, University of Oxford, Oxford, UK
e-mail: simon.saunders@philosophy.ox.ac.uk

© Springer Nature Switzerland AG 2020 479

M. Hemmo, O. Shenker (eds.), Quantum, Probability, Logic,
Jerusalem Studies in Philosophy and History of Science,
https://doi.org/10.1007/978-3-030-34316-3_22
480 S. Saunders

(it was by then accurately confirmed at high frequencies and low temperatures).
The derivation was what John Norton has called Einstein’s ‘miraculous argument’,
contained in the most revolutionary of all of Einstein’s great papers of 1905.3
If it was derived from the data, how was it opposed? The answer in part is that it
was incomplete: whatever light was really made of, it could not only be Einstein’s
light quanta, because if it were, it would contravene the Planck distribution in the
low frequency high-temperature regime (the Rayleigh-Jeans regime, also by then
accurately confirmed). What was the nature of the struggle, that was so intense
and prolonged? The answer in part was the difficulty of reconciling a particulate
structure to radiation with Maxwell’s equations (or any wave equation, needed to
explain interference effects), but another part was down to Einstein’s insistence
on statistical independence. It was Einstein’s struggle; and it was only concluded
(and then only temporarily) with his quantum theory of the ideal monatomic gas,
published in three instalments in the period 1924–1925.
To see the nature of the difficulty, consider that radiation in thermal equilibrium
behaves as a gas of indistinguishable particles, photons, satisfying the Bose-Einstein
statistics; and that the latter, as Einstein admitted in his second paper of his gas
trilogy, implies that particles are not statistically independent of each other; yet
Einstein’s ‘miraculous argument’ of 1905 had shown that thermal radiation, in the
Wien regime, is made of statistically independent particles. Contradiction.
An easy way out is to suppose that in the Wien regime photons just are
statistically independent of each other; and it is true that in this regime they satisfy
Maxwell-Boltzmann statistics. However, the latter is not the signature of statistical
independence: classical indistinguishable particles, in Gibbs sense, obey Maxwell-
Boltzmann statistics, but they are not statistically independent in the requisite sense.
The better answer is that Einstein did not show that radiation, in the Wien
regime, is composed of statistically independent particles. His argument showed
that in the Wien regime thermal radiation is made of non-interacting, localised
entities; it left open the question of whether they were indistinguishable, in Gibbs’
sense, or statistically independent, in Einstein’s. However, as we shall see, a slightly
different kind of fluctuation than the one Einstein considered resolves the under-
determination in favour of the Gibbs concept – for roughly the reasons Gibbs
himself advocated it, namely, as a solution to what is now called the Gibbs paradox.
Had Einstein in 1905 understood this point, the history of the development of
quantum theory would have been significantly different, for it would have been a
relatively short step to Einstein’s quantum gas theory. It is therefore something of
an embarrassment that Gibbs’ concept of indistinguishability, introduced in terms
of the notion of ‘generic phase’ (as opposed to ‘specific phase’, as had hitherto been
implicitly assumed), was to be found in his Elementary Principles in Statistical
Mechanics, published in 1902, and indeed already published in German translation
in 1904.

3 Norton (2006).
22 What Price Statistical Independence? How Einstein Missed the Photon 481

It seems that Einstein knew nothing of Gibbs’ work at this time, but we know
that he had read at least parts of the Principles by 1910 (and that it impressed him
greatly)4 ; however, we also know that he was sceptical of Planck’s claims, in the
period 1912–1918, on the extensivity of the entropy function of a material gas, that
cited Gibbs5 (the failure of extensivity is one formulation of the Gibbs paradox).
When Einstein embraced essentially the same solution to the extensivity puzzle in
the second of his papers on gas theory, it was in the name of what he called ‘Bose’s
statistical approach’, not Gibbs’.
Bose himself, by his own admission, was unaware that his approach had differed
from the usual one – and, indeed, his procedure so obscured the underlying
statistical assumptions that Einstein, when he took it over in his first paper on gas
theory, did not realise (or at any rate did not mention) that it implied the failure of
statistical independence. In his second paper, published at the beginning of 1925,
Einstein presented the expression for the entropy in a different form, from which
this failure was ‘easily recognised’; but he still attributed the approach to Bose.
That expression for the entropy, I am claiming, is the natural transposition of Gibbs’
notion of generic phase to the case where the one-particle state space has a finite
number of states (when it is divided into ‘elementary cells’, of size h3 ) – the central
assumption of the ‘old’ quantum theory.
No more did Schrödinger identify Gibbs as the progenitor of ‘Bose’s approach’,
although he came close. He had publicly criticised Planck ‘s arguments for an
extensive entropy, that drew on Gibbs’, and now found fault with Einstein’s new gas
theory precisely on the point that, following Bose’s method, particles could not be
statistically independent of each other. Einstein, by then, was prepared to grant that
it was for nature to decide; not so Schrödinger, who went on (in a paper immediately
preceding the first of his historic papers on wave mechanics) to propose a new
quantisation method for the quantum gas that maintained statistical independance.
Neither Einstein nor Schrödinger mentioned Gibbs by name, nor the concept of
generic phase. The following year, in 1926, Dirac laid down the definitive treatment
of particle indistinguishability in quantum mechanics, again with no reference to
Gibbs (or to Schrödinger on gas theory) – and few physicists looked back to its
precursors since.
Historians, meanwhile, seem to have taken Einstein, Schrödinger, and Dirac
rather too much at face value; since none of them mentioned Gibbs or the generic
phase, the equivalence of the latter with ‘Bose’s approach’ has not been recognised,
or even seriously entertained.
There are other reasons why the equivalence was not recognised. One is because
Bose-Einstein statistics is directly related to the symmetrisation of the state in
quantum mechanics (Dirac cited Bose and Einstein). But the latter, by wide
consensus, is thought to involve entanglement, a quintessentially quantum concept.
That it is in fact entanglement of a particularly trivial kind – that does not, for

4 As recounted by Pais (1982, p. 55).

5 From correspondence with Ehrenfest in 1918, quoted in Darrigol (1991, pp. 288–89).
482 S. Saunders

example, support the violation of any Bell inequality – has only recently been
shown. Another reason the equivalence has not been recognized comes from the
classical side, where it is thought that the concept of particle indistinguishability
cannot apply to classical mechanics ‘because classical particles can always be
distinguished by their trajectories’. That phrase, or one like it, recurs again and
again in the subsequent literature.6 But it meets the counter-argument that quantum
particles can also have trajectories – either approximately, for sufficiently massive
and weakly-interacting particles in well-localised states, or exactly, in the case of
non-interacting particles, taking ‘trajectories’ as orbits of orthogonal one-particle
states. In neither case do the particles cease to be indistinguishable, in the sense
that their total state remains symmetrised7 We shall see a simple example in the
concluding section.
It appears that the struggle initiated by Einstein with the photon concept is so
intense, and so prolonged, that it continues unabated to this day. I can offer up
nothing quite as remarkable as Itamar Pitowsky’s discovery, in George Boole’s
‘impossibilities of experience’, of Bell’s inequalities. His beautiful theory of corre-
lation polytopes has long been a source of inspiration to me, as has his independence
of thought and temperament. But it is my hope that he would have found pleasing
the discovery in Gibbs’ paradox of the kind of failure of statistical independence
on which quantum indistinguishability is based – as within the realm of possible
experience. Jon Dorling, in his acute historical and theoretical observations, was a
similar inspiration, and himself wrote on my topic, arguing that in the limit of the
Wein regime Einstein’s argument showed light quanta were truly point-like. Their
friendship and support has always meant much to me. This essay is dedicated to
Itamar and to Jon.

22.2 The Gibbs Paradox and ‘Mutual Independence’

Boltzmann’s equation

S = k ln W + const

was first written down in this notation by Planck, but it was introduced by Boltzmann
in his 1877 memoire, often accounted his most important contribution to statistical
mechanics. It was derived anew by Einstein in 1905 on the basis of what he
called ‘Boltzmann’s principle’, namely the principle that the entropy of a system

6 See,for example, Epstein (1936a, p. 557), Darrigol (1986, pp. 205–9), Bach (1997, p. 8), Dieks
(2014). (Bach’s defence of the concept of classical particle indistinguishability extends only to the
notion of probability distributions that give no information on individual trajectories.)
7 Saunders (2006, pp. 199–200, 207).
22 What Price Statistical Independence? How Einstein Missed the Photon 483

is a function of the probability W of its instantaneous state. His argument was as

follows8 :
If it makes sense to talk about the probability of a state of his system, and if, further, each
entropy increase can be conceived as a transition to a more probable state, then the entropy
SA of a system A is a function of the probability WA of its instantaneous state. Therefore, if
we have two systems SA and SB that do not interact with each other, we can put

SA = ϕ A (WA ) , SB = ϕ B (WB ) .

If these two systems are viewed as a single system of entropy S and probability W, we
have

S = SA + SB = ϕ(W ) (22.1)

and

W = WA .WB.

This last relation tells us that the states of the two systems are mutually independent
events.
From these equations it follows that

ϕ (WA .WB ) = ϕ A (WA ) , ϕ B (WB ) ,

and from this we get, finally,

ϕ A (WA ) = C ln (WA ) + const

ϕ A (WA ) = C ln (WA ) + const

ϕ (W ) = C ln(W ) + const.

The quantity C is thus a universal constant; it follows from the kinetic theory of gases
that its value is R/N [Boltzmann’s constant k, in Planck’s notation].

The argument has power and beauty, but we must look to its premises. The
left-hand equality of (22.1) is nowadays called the additivity of the entropy. In
thermodynamics, it applies to non-interacting systems A and B each in thermal
equilibrium (not necessarily with each other), but Einstein went on to use it to
broader effect. Let there be z one-particle states, all equiprobable, so that the
probability that a particle will be found in a given state is 1/z, and suppose there are n
non-interacting particles. They will be ‘mutually independent’ in Einstein’s sense if
the probability that they will all be found in that state is 1/zn . The joint probability, in

8 Einstein (1905, p. 94–95); I follow Einstein’s notation, save for replacing ‘1’ and ‘2’ by ‘A’ and
‘B’.
484 S. Saunders

other words, should factorise, as the product of probabilities for individual particles
considered separately. In this sense, the probability of a set of outcomes on throwing
of n dice should be the same, whether the dice are thrown sequentially or all at once,
so long as they are thrown independently and do not influence each other.
If thermodynamic probabilities are to factorise in this way when particles are
non-interacting, and if they are to connect in a simple way with the available phase
space volume,9 then the volume measure on the many-particle phase space had
better factorise; and that in turn requires that the phase space itself factorises, i.e.
that it is the Cartesian product of 1-particle phase-spaces γ:

Γ = γ × .... × γ (22.2)

with volume measure:

WV = V n .

The dependence of the entropy of n particles on the spatial volume V is then of

the form

SV = k ln V n . (22.3)

There is the inevitable attendant failure of extensivity: SV as given by (22.3)

does not scale with the extensive quantities, n and V. Double the latter and the
entropy is not only doubled, but picks up an additional term 2nk ln 2. Of course
it is true that extensivity is something of an approximation, as the molecules making
up real material gases are at least weakly interacting and there may be boundary
effects, but it is an extremely good approximation for thermal radiation, so long as
the temperature is not too high, and it follows, or should follow, for an ideal material
gas as well.
In terms of the set-up of the Gibbs paradox, suppose two samples of the same
ideal gas are confined to volume VA (nA particles) and VB (nB particles), at equal
pressure and temperature. The initial total entropy is then (from additivity):

SA+B = SA + SB = k ln VAnA + k ln VBnB . (22.4)

Let the two volumes be contiguous and let the two gases subsequently diffuse
into each other. Then the final entropy will be:

SA∪B = k ln (VA + VB )(nA +nB ) . (22.5)

9I shall use the terms ‘probability’ and ‘phase space volume’ more or less interchangeably, and
lazily use the same symbol V for both spatial volume and volume of γ; the context should make
clear which is intended.
22 What Price Statistical Independence? How Einstein Missed the Photon 485

The result is an increase in entropy; in the simplest case VA = VB the difference

between (22.4) and (22.5) is 2nk ln 2. It is the same term spoiling extensivity that we
encountered before, where now it reflects the two possibilities for each particle after
the partition is removed, whether located in A or in B, where before there was only
one. This is the form of the Gibbs paradox that we shall be concerned with, where
there is no exchange of particles with outside systems, but only between particles in
the two volumes VA and VB .
In order to obtain no entropy of mixing, and thereby an extensive entropy
function, Gibbs proposed to use the quotient of  as given by (22.2) under the
permutation group instead,10 to use the space of what he called generic phases
(now usually called reduced phase space):
∼
Γ = γ × . . . . × γ/Π.

It is the space arrived at by identification of points in  that are related by a

permutation of particles (a permutation with respect to which particle has which
position and momentum; a permutation of sequence-position in the Cartesian
product). The phase space (22.2) was called by Gibbs the space of specific phases.
As Gibbs’ remarked, the reduced phase space has a considerably more complicated
topology, but the volume measure has a simple expression:

∼ Vn
WV = . (22.6)
n!
The latter does not factorise, and particles described in this way fail Einstein’s
test of statistical independence. For suppose particles nA are statistically indepen-
dent from particles nB , both occupying the same region V. Then the probability
∼
of the instantaneous state of the composite W A∪B should be the product of the
∼ ∼
probabilities of the state of each, W A ·W B . However, by inspection:

∼ V (nA +nB ) V nA V nB ∼ ∼
W A∪B = ≈ · = WA · WB (22.7)
(nA + nB )! nA ! nB !

and the probability does not factorise. Contrast the specific probabilities, that do:

WA∪B = V (nA +nB ) = V nA · V nB = WA · WB .

Returning to the generic phase, our argument used ‘A’ and ‘B’ as names for
the two collections of particles, dispersed through the same volume, when they
are supposed to be indistinguishable; is this move legitimate? But surely it is if

10 Gibbs did not of course use the concept ‘group’, and gave no independent motivation for the
generic phase other than that it gave no entropy of mixing. It was seen as ad hoc by Lorentz and
Ehrenfest; see Saunders (2020), Darrigol (1991) for further discussion.
486 S. Saunders

‘indistinguishable’ is merely to renounce factor position in the Cartesian product

Eq. (22.3) – what I shall call the realist reading of the Gibbs’ generic phase. The
n-tuple of pairs in a definite sequence:

qa , pa , qb , pb , . . . , qd , pd   ∈ Γ

where qa , pa  ∈ γ , etc., is a specific phase (specific phase space point). The
identification of all such points related by a permutation of entries in sequence
position, is to pass to a space the points of which are represented by unordered
collections:
∼
{qa , pa , qb , pb , . . . , qd , pd } ∈ Γ .

There is no longer a sequence position to pick out a particle, independent of

its position and momentum, but the latter can be specified as such directly, and
thereby the particle that has that position and momentum, supposing there are no
repetitions.11 The identification of points in  related by permutations does not
∼
mean that the resulting space  does not have points. Nor does it mean that we must
interpret those points as an equivalence class of specific phases (sequences); but as
Gibbs remarked, the latter is useful from the point of doing analysis on manifolds.12
There is nevertheless a fundamental distinction between this way of specifying
the partition of the gas of n particles into two collections, each scattered throughout
the same spatial volume, and specifying the partition in terms of disjoint volumes
instead (and indeed disjoint regions of the one-particle phase space, regions in both
momentum space and position space). Specify the sub-systems A and B in the
latter way, and we are back to the Gibbs set-up. The generic probability now does
factorise13 :
∼ ∼ ∼ V nA V nB
W A+B = W A .W B = · .
nA ! nB !

If the two gases are allowed to mix, they are no longer statistically independent
of each other, as we have seen, and the joint probability is given by the left-hand
side of Eq. (22.7). Using the Sterling approximation:

11 The assumption is innocent in a continuum theory of point-like particles, or for extended but
impenetrable molecules, but for a finite state space it is the difference between fermions (no
repetitions) and bosons (repetitions).
12 Gibbs (1902, p. 188). Whether my realist reading of the generic phase should be attributed to

Gibbs is a delicate question, for Gibbs writings were so brief. For further discussion, see my (2020).
13 If this is in doubt, consider the analogous question using a discrete state-space, and apply Eq.

(22.16) to find the number of ways of distributing nA indistinguishable particles over zA states, and
likewise for nB particles over zB disjoint states; the product of the two numbers is the same as the
number of ways of distributing n = nA + nB indistinguishable particles over z = zA + zB states,
under the constraint that nA are in VA and nB are in VB (with no regard to which is in which).
22 What Price Statistical Independence? How Einstein Missed the Photon 487

ln n! ≈ n ln n − n

and bearing in mind that since the two gases prior to mixing had the same
temperature and pressure, their densities are equal as well, nA /VA = nb /VB ; it follows
after a short calculation:
∼ ∼
W A+B ≈ W A∪B . (22.8)

The reason the equality is only approximate in (22.8) is because after the gases
mix the states available include those in which numbers of particles different from
nA and nB may be found in the regions A and B respectively (but not of course
by virtue of which particles are found in each of the two regions). After mixing,
the available number of microstates is thus increased, but by a number that can be
neglected in the Sterling approximation.
The lesson of the Gibbs paradox is that indistinguishable particles, in Gibbs’
sense, even if non-interacting, are not statistically independent in Einstein’s sense.
To put the matter vividly: if there are nA particles alone, initially confined to volume
VA , expanding into volume V, the entropy increases. If there are nB particles alone,
initially confined to volume VB , with the same density, expanding into volume
V, the entropy increases. Yet let nA , nB particles be present in each of the two
volumes, of the same density, and allowed to expand into volume V, then there is no
entropy increase, even though the particles are non-interacting, and even though the
trajectory of every particle is the same as if all the others were not there. The entropy
of the joint system is not the sum of the entropies of the two subsystems, and the
entropy change of the joint system is not the sum of the entropy changes of the two
subsystems. Einstein’s criterion fails for both, and with it mutual independence.

22.3 Einstein’s ‘Miraculous Argument’

Consider now Einstein’s celebrated argument from the Wien black-body distribution
formula. This claimed to show that insofar as that equation is obeyed, light behaves
as if made of elements that were ‘mutually independent’ and, moreover, that were
‘moveable points’.
First ‘mutually independent’. We must take Einstein to mean what he meant
earlier in the same paper, in the argument for the universality of the entropy as a
function of W just reviewed. Mutual independence, then, means the factorising of
probabilities. Is it the same as the condition that the particles are non-interacting? It
has been interpreted in this way,14 but that would be a mistake: even if Einstein
did ultimately believe that statistical correlations must be explained by some

14 Most recently by Norton (2006), according to whom ‘independence is just the simplest case of
no inter-molecular forces’ (p. 93).
488 S. Saunders

dynamical coupling, he consistently used the phrase ‘mutual independence’ and

later ‘statistical independence’ (and not ‘non-interacting’).
Given all this, and given that the Wien limit to the radiation law just is the
limit in which Gibbs’ measure for the generic phase is valid, there is the obvious
difficulty: the probabilities as defined by Gibbs’ measure do not factorise, as we
have just seen. How then, working from the Wien regime, could Einstein have
concluded the probabilities factorise, and hence that light quanta were mutually
independent of each other? Indeed, on a careful reading, this was not exactly
his claim, nor an accurate reading of his argument. He actually showed that if
thermal radiation consisted of mutually independent and localised entities, then
its spectral distribution would be the Wien distribution. From this we infer that
thermal radiation, in the Wien limit, behaves as if made of mutually independent
and localised entities, not that it is made of such entities.
Einstein himself, in his summary conclusion, put the onus entirely on mutual
independence:
Monochromatic radiation of low density (within the range of validity of Wien’s radiation
formula) behaves thermodynamically as if it consisted of mutually independent energy
quanta of magnitude hv.

The omission of locality may have been no more than a slip; does he commit the
fallacy of affirming the consequent as well? He treads a fine line: light behaves as
if it is so constituted, granted; but it so behaves if it is constituted in other ways as
well – in particular, if the energy quanta are non-interacting and localised but lack
statistical independence, in the very particular way that they are correlated when
described by the generic phase.
We clearly need to see the details. From the first law of thermodynamics:

dU = T dS + pdV

it follows:

dS  1
 = .
dU V =const T

Einstein worked directly with the spectral entropy density ϕ of black-body

radiation (the entropy per unit frequency range and volume), as a function of the
spectral energy density ρ (the energy per unit frequency and volume), to derive
what he called ‘the law of black bodies’:
∂ϕ 1
= . (22.9)
∂ρ T

The Wien spectral distribution law, as first proposed by Wilhelm Wien in 1896,
is

ρ = αv 3 e−βv/T (22.10)
22 What Price Statistical Independence? How Einstein Missed the Photon 489

where α and β were new and unknown constants. It was well-confirmed by

experiment for v/T ! 1, as Einstein noted. Using (22.10) to write the inverse
temperature as a function of ρ and the frequency, and inserting in (22.9), Einstein
obtained a differential equation for ϕ. It could readily be integrated to obtain:
ρ  ρ 
ϕ=− ln 3 − 1 . (22.11)
βv αv

Planck had found this equation in 1898, likewise working from the Wien
distribution (and he had used the same method again in 1900, to obtain an equation
for the entropy as a function of ρ, working from the Planck distribution). Einstein’s
originality lay in what came next.
Let S0 = V0 ϕ be the entropy of the radiation in volume V0 per unit frequency
range and let E be the energy V0 ρ. Multiplying both sides of (22.11) by V0 , it
follows, writing ρ in terms of E and V0 :
 
E E
S0 = − ln − 1 .
βv αv 3 V0

Suppose now that this volume of radiation (in the given frequency range)
undergoes a spontaneous fluctuation, and is contained instead in the sub-volume
V; then it will have the entropy:
 
E E
S=− ln 3 − 1 .
βv αv V

The entropy change is therefore:

E V
S − S0 = ln
βv V0

or more suggestively:
  E
V kβv
S − S0 = k ln . (22.12)
V0

It is the same volume-dependence of the entropy function as for an ideal gas

of n = E/kβv particles; in Planck’s notation, for n = E/hv particles; in Einstein’s
words, as for a gas of n independently movable points.
Here is Einstein’s argument in full:
Let us consider a part of the volume V0 of magnitude V and let all n moveable points be
transferred into the volume V without any other change in the system. It is obvious that
this state has a different value of entropy (S), and we now wish to determine the entropy
difference with the aid of Boltzmann’s principle.
490 S. Saunders

We ask: How great is the probability of the last-mentioned state relative to the original one?
Or: How great is the probability that at a randomly chosen instant of time all n independently
movable points in a given volume V0 will be contained (by chance) in volume V?
Obviously, for this probability, which is a ‘statistical probability’,15 one obtains the value16
* +n
V
W rel = ; (22.13)
V0

from this, by applying Boltzmann’s principle, one obtains

 n
V
S − S0 = k ln . (22.14)
V0

By ‘probability . . . .relative to the original’ we take it Einstein means the ratio of

the probabilities of the two states, so that, in accordance with (22.14), its logarithm is
the difference in the entropies of the two states.17 And whether or not (22.13) really
is obvious: as before, we take Einstein to mean that if independent, the probability
for the joint state of all n moveable points is the product of the probabilities for each;
or at least that this is true of the relative probabilities. The relative probability that a
single particle, initially in volume V0 , is subsequently in the sub-volume V, is V/V0 ;
hence the relative probability that this is true of all n particles is the product of the
relative probabilities of each, Eq. (22.13).
This is, ‘obviously’, consistent with the volume dependence of the probability on
the total phase space of the n particles of the form already considered:

W = V n.

It reproduces the required relative probability, in that

 n
W V
W rel
= = .
W0 V0

It is the same as the ratio of probabilities for Gibbs’ specific phase. But the ratio
of probabilities for Gibbs’ generic phases is exactly the same:
∼  n
∼ W V
W rel
= ∼ =
V0
W0
∼
even though W , as given by Eq. (22.6), cannot be factorised. Non-interacting move-
able points that are indistinguishable in Gibbs sense are not mutually independent in

15 The reference is to his preference for an ergodic interpretation of probabilities, rather than
Boltzmann’s combinatorial method, that he had emphasised in his early papers of 1903 and 1904.
16 To avoid confusion, we write Wrel where Einstein wrote W.
17 This is consonant with e.g. Pais (1982, p. 72); I am not aware of any alternative reading.
22 What Price Statistical Independence? How Einstein Missed the Photon 491

Einstein’s sense, yet yield the same relative probability and have the same volume
dependence of their entropy as do mutually independent movable points. The Wien
law, concerning this kind of fluctuation, does not discriminate between them.
Would it be better to read Einstein’s notion of independence as requiring the
factorising only of relative probabilities? In that case indistinguishable particles
in Gibbs’ sense may come out as statistically independent after all. But that
contravenes what we have learned from Gibbs’ paradox in Sect. 22.2. Indeed,
consider a fluctuation of a rather different sort than the one Einstein imagined,
in which nA particles initially occupying volume V are subsequently found in
the sub-volume VA , and the remaining nB particles, initially in volume V, are
subsequently found in the disjoint sub-volume VB , where V = VA + VB . Suppose
first that the nA particles are statistically independent of the nB particles. Then the
relative probability of the fluctuation should factorise into the product of the relative
probability of each considered independently:
 nA  nB
VA VB
W rel = WArel .WBrel = . .
V V

It is the same, whether calculated as the ratio of the probability of the joint final
state to the initial, or as the product of the ratios of final and initial states of each
separately. Suppose now that the fluctuation does not change the density, so that:

(nA + nB ) nA nB
≈ ≈ .
(VA + VB ) VA VB

The fluctuation is then the time reverse of a process of mixing of gases, as in

the Gibbs set-up. It follows there is a decrease in entropy (for VA = VB , a decrease
equal to nk ln 2), which is as to be expected, given that the reverse process yields an
increase in entropy.
If now we use generic phases instead, there is no such decrease, for the same
reason that the time-reverse process gives no increase. The relative probability as
given by the ratio of the probability of the joint final state to the probability of the
joint initial state is:

∼ VA nA /nA ! · VB nB /nB !
W rel =
V N /n!

which, as we have already seen, is approximately unity in the Sterling approxima-

tion. It is not the product of the relative probability for a fluctuation in the state of nA
particles, initially occupying volume V, by which they are found in volume VA , with
the relative probability for a fluctuation in the state of the remaining nB particles,
from V to VB . Eq. (22.7) holds here just as it did for the Gibbs paradox; therefore
the relative probability does not factorise.
Our earlier conclusion therefore stands: indistinguishable classical particles in
Gibbs’ sense are not statistically independent in Einstein’s sense, not even if
492 S. Saunders

Einstein’s factorizability condition is only required for relative probabilities. Since

photons in the Wien limit have the same functional dependence on the volume and
particle number as do indistinguishable classical particles, in Gibbs’ sense, the result
applies as much to photons. Photons are not statistically independent, in Einstein’s
sense, not even in the Wien regime.

22.4 Gibbs’ Generic Phase as the Limit of Bose-Einstein

Statistics

There are of course other solutions, or purported solutions, to the Gibbs paradox, or
rather paradoxes (for it has been taken to mean many different things), than the one
we have considered; and there are other criticisms, or purported criticisms, of Gibbs’
concept of generic phase. But there is a clear structural correspondence between
statistics as made out in terms of the generic phase, vs the specific phase, and
statistics following what Einstein in 1924 called ‘Bose’s statistical approach’, vs that
for ‘statistically independent molecules’, following Boltzmann’s. Bose’s method, in
the Wien limit, is the same as Gibbs’, using the generic phase.
Alas, neither this structural correspondence, nor the connection with Gibbs
paradox, was so much as mentioned by any of the key players, nor were they for
some years.18 The first to consider the matter in any detail appears to have been
Paul Epstein, in a contribution to the Commentary on the Scientific Writings of J.
Willard Gibbs published by Yale University, Gibbs’ alma mater, in 1936. Epstein
highlighted the relationship between specific vs generic phase, in Gibbs’ sense, and
statistical independence vs Bose’s method, in Einstein’s; but he also maintained,
as did virtually everyone following Dirac’s 1926 intervention, that the absence of
trajectories plays an essential role in quantum indistinguishability, which therefore
has no analogue in classical statistical mechanics (whether using specific or generic
phases)19 – precisely the opposite conclusion to the one advanced here.
Epstein’s intervention, whatever its merits, came too late for historians of the
indistinguishability concept. It came too late for Gibbs, for by then photons were
a species of bosons; gibbsons were not to be. Yet the connection was there for
the taking in Einstein’s discussion. The key expression for the thermodynamic
probability, the signature of the new statistics, had been at the heart of the quantum
enigma since it first appeared in Planck’s publication of 1900 (where Planck’s
symbol ‘h’ appeared for the first time as well):

18 As remarked, Schrödinger (1925) came the closest, but did not mention Gibbs or the generic
phase. For further reading, see Darrigol (1991, pp. 293–8), Mehra and Rechenberg (1987, pp.
361–66), Saunders (2020).
19 Epstein (1936a, p. 557). Epstein also argued that a discrete state space was inconsistent with

a continuous dynamics, in the case of generic phase, in (1936a, p. 564, p. 527) and (1936b, pp.
485–91).
22 What Price Statistical Independence? How Einstein Missed the Photon 493

(n + z − 1)!
. (22.15)
(z − 1)!n!

It has a simple combinatorial meaning: it is the number of ways of distributing n

objects, among z cells, where it makes no difference as to which object is assigned
each cell, but only how many. It is the count of microstates, where each is defined
as a specification, of each cell, of the number of objects it contains, but not of which
of the n objects is assigned each cell. This was well-understood by 1911 if not
already by Boltzmann in 1877 (who was the first to write down (22.15), as the
number of ways that ‘energy elements’, meaning integral multiples of a unit ε could
be distributed among n objects). It may well have been known to Gibbs, although
Gibbs probably had no reason to take seriously the idea that z might have a well-
defined physical meaning (for that was provided by Planck’s discovery in 1900,
putting ε = hv, and as later developed in terms of the number of elementary cells to
phase space of volume h3 ). If Gibbs read Planck’s publications in 1900 and 1901,
he did not comment on them; he died unexpectedly in 1903.
Planck’s notation did not make evident that the numbers n and z were tied to a
range of frequency and position, denote [s, s + Δs]. Einstein’s notation, 25 years
later, did, and following Bose, generalised it from a range of frequency to a
range of momentum (as well as spatial volume). Such a range Einstein called an
elementary region s of the 1-particle state space, of the form V × [p, p + Δp] ⊂ γ
(provided by what we would today call a coarse-graining of the one-particle state-
space). Dividing by h3 gave the number of elementary cells zs . For the number of
microstates, if there are ns particles distributed over the zs cells of region s, and if a
microstate only says how many are in each cell (‘Bose’s method’), Einstein wrote
down the expression:

(zs + ns − 1)!
(22.16)
(zs − 1)!ns !

immediately remarking that ‘it is easily recognized’ that the distribution of

∼
molecules over cells is not statistically independent. We shall denote it Z s , as
∼
the analog of W s , the continuous measure on the reduced phase space of the region
s. The total probability of a macrostate is then the total number of microstates,
consistent with the specification of the numbers ns , for all the regions s, i.e. the
product20

∼ % (zs + ns − 1)!
Z= . (22.17)
s
(zs − 1)!ns !

20 The probability factorises for the elementary regions, in the same way and for the same reasons,
as it does for the subsystems defined by VA and VB in the Gibbs set-up (see fn.13).
494 S. Saunders

Einstein contrasted this expression with that following from ‘the hypothesis of
statistical independence of the molecules’. There ‘a state is defined microscopically
by the fact that for each molecule one indicates in which cell it is sitting (complex-
ion)’,21 where ‘complexion’ was Boltzmann’s 1877 term for microstates. In that
case the number of microstates of ns particles in region s is:

zs n s ,

denote Zs , and across all the regions for given occupation numbers n1 , n2 . . . it is

n! %
zs n s (22.18)
n1 !n2 ! . . . . . . s

where n is the total number of particles, and where the pre-factor

n!
n1 !n2 ! . . . . . .

is Boltzmann’s famous ‘permutability’, as he called it. It reflected the additional

multiplicity in microstates, introduced when partitioning the total number of
molecules n among the elementary regions, so that n1 are in region 1, n2 in region
2, and so on.
The different expressions (22.17) and (22.18) yield different entropy functions.
Only one can be correct – and on this point Einstein thought the question was an
empirical question, that ‘nature must decide’. But of course, in the case of thermal
radiation, nature’s decision was already known, favouring the Bose-Einstein entropy
and the new statistics.
And there, more or less, Einstein left the matter. He did not comment on the
Wien limit; neither did Bose (and nor, in 1926, did Dirac). We should verify that it
is indeed the limit of the Gibbs generic phase. The Planck distribution is:

8π v 2 hv
ρ= .
c 3 e hv/kT −1

It goes over to the Wien distribution (22.10) in the limit hv ! kT. But that is
the limit in which very few modes of the electromagnetic field are excited, for the
minimum required is much greater than the thermal energy available (of order kT),
and those that are excited are at the lowest excitation (so the occupation numbers
are all 0’s and 1’s). Since zs is the number of modes, it follows the Wien limit is the
limit zs ! ns . In that limit it can easily be seen:

21 Einstein
(1925); ‘each molecule’ appears to be functioning as a proper name. See Saunders
(2006, 2013, 2016) for more on philosophy of language and indistinguishability.
22 What Price Statistical Independence? How Einstein Missed the Photon 495

(zs + ns − 1)! zs n s
≈ . (22.19)
(zs − 1)!ns ! ns !

The total number of microstates for all elementary regions s for which zs ! ns is
then approximately:
% zs n s
(22.20)
ns !
s;zs !ns

recovering the Maxwell-Boltzmann probability Eq. (22.18), save for the overall
factor n! (where now n is the total number of particles in the Wien regime). In
∼ ∼
terms of the continuous phase space measure W s for ns particles in region s, Z s and
∼
W s are proportional only in the Wien regime. In contrast the equation

1
Zs = Ws
τ ns
holds for every region s and numbers ns , where τ is the unit of phase space volume
γ . Classically τ can always be chosen as sufficiently small to ensure the Wien limit
is enforced (only the size of the elementary regions must be kept large, so as to
contain many particles, so that the Sterling approximation applies). Boltzmann’s
combinatorial method then goes over unchanged (apart from the overall factor
n!), using Gibbs’ generic phase, on replacing ‘cells’ by ‘elementary regions’, the
same as using the specific phase. In the quantum case τ = h3 , where h is Planck’s
constant, whereupon the number of cells zs cannot be taken as arbitrarily large, and
the approximation (22.19) fails for sufficiently large ns .
None of this was mentioned by Einstein. The correspondence between (22.18
and 22.20) should have been obvious much earlier, for example, to the Ehrenfests,
at the time of writing their famous Encyclopaedia article in 1912, for they knew of
the distinction between elementary regions and cells (which indeed goes back to
Boltzmann in 1877). They too did not mention it.22
Fast-forwarding to Dirac, all of the foregoing applies to equilibrium distributions
of photons. In the Wien regime, the total number of microstates in each elementary
region is well-approximated by (22.19), so photons are not statistically independent,
in Einstein’s sense. The exact expression for every frequency and temperature (so
long as the temperature is not so high that pair-creation of electrons becomes
∼ ∼
possible) is Z s , as given by (22.16). Indeed, Z s is exactly the dimension of the
subspace of the Hilbert space spanned by symmetrised wave-functions in the ns -fold

22 Theyassigned zs = ω in Eq. (22.18), with ω an arbitrary constant independent of s; whereupon

the comparison of (22.20) with (22.18) is all but invisible (Ehrenfest and Ehrenfest 1912, p. 27).
Ehrenfest and Kamerlingh-Onnes in their (1914) came very close to the identification, but also
missed it; see my (2020) for further discussion.
496 S. Saunders

tensor product of the one-particle Hilbert space Hs , corresponding to the range of

frequency (momentum) [s, s + Δs] in spatial volume Vs , where Hs has dimension zs .

22.5 Locality and Entanglement

If Einstein was wrong to think that in the Wien limit light was made up of
statistically independent entities, his argument surely did show that in that regime
it has a discrete structure, and consists of non-interacting localised entities23 – in
that nobody has been able to come up with another interpretation of the fluctuation
formula (22.12) that says any different. He was right in another respect as well:
in recommending that light quanta be thought of as movable. That might seem
an obvious presupposition of his thought-experiment (how else was the fluctuation
in volume supposed to happen unless the particles were ‘movable’?), but it was a
first step towards investing the photon, as the concept was eventually named, with
different properties at different times – in short, to considering it as the sort of
entity that might have different states at different times, and hence that might have a
state-space at all; whereas ‘energy elements’, of the sort that had been discussed by
Boltzmann, Planck, and Natanson, were never thought of like that. The latter were
alike in all of their properties; they were hardly thought of as capable of change
at all, unless it be the transfer of energy, emission and absorption, creation and
annihilation.
Did it matter, historically, that light quanta were wrongly thought to be statisti-
cally independent in the Wien regime? For those in the critical period 1905–1925
prepared to extend the concept of light quanta to the Rayleigh-Jeans regime, it led
to the idea that correlations only came into play in that regime. This encouraged the
picture of a cooperative behaviour (‘the light molecule’) of quanta all in the same
cell, or all in the same normal mode of cavity radiation. Einstein was sympathetic
to these ideas. They were also entertained, in one of his early papers, by Louis de
Broglie. Einstein warmed to this theme in his second paper on gas theory in 1925.
He recommended de Broglie’s newly published PhD thesis, as taken up, among
others, by Schrödinger. In these respects the confusion was positively beneficial,
for it led Schrödinger to de Broglie’s concept of matter wave and to his concept of
quantisation conditions as defined by solutions to wave equations; but confusion it
was.
The story did not stop there, however. Einstein called the failure of statistical
independence ‘an initially completely puzzling mutual influence of the molecules’
(ibid, p.374). He also spoke of ‘spooky’ action-at-a-distance, a line of inquiry that
eventually led to the EPR paradox. The latter, we know, exploits entanglement.

23 Whether the ‘movable points’ must really be points will depend, presumably, on the precision
with which the probability of the fluctuation is defined (perhaps only in the limit v → ∞). I refer
to Dorling (1971) for discussion, already mentioned.
22 What Price Statistical Independence? How Einstein Missed the Photon 497

(Anti-)symmetrisation of the wave-function also involves entanglement, and surely

explains the failure of statistical independence. The two, it might be thought, are at
bottom the same, an argument developed at length by Don Howard (and recently
endorsed by Norton).24 Yet there are important differences: as already remarked,
there is a growing body of literature to show that the particular kind of entanglement
required by symmetrisation is essentially trivial. Not only is it consistent with the
existence of well-defined particle trajectories, it is also insufficient to violate any
Bell inequality.25 It is, I suggest, just what is required, no more and no less, to ensure
that factor-position, on taking the tensor-product of one-particle Hilbert spaces, can
play no role in the definition of any physical quantity. Thus the partial trace of the
state, as defined by factor-position, with respect to one or more of all other particles,
delivers exactly the same density matrix whatever the factor-position (a result that
has led some madmen to conclude that every electron is in exactly the same state).
Then what does explain the failure of statistical independence? An alternative
but still somewhat tentative answer is that microstates should not be considered as
built up out of particle motions taken one at a time; that a microstate at an instant of
time supervenes globally, rather than locally, on the states of individual particles.
The idea that the dice thrown all at once should yield outcomes with the same
probabilities, if non-interacting, as were each die thrown sequentially in time, may
be the culprit; for if n dice are thrown sequentially in time, each with z equiprobable
outcomes, there will be

n!
n1 !n2 ! . . . nz !

different possible temporal sequences, all with the same final set of occupation
numbers n1 , . . nz , with each sequence equiprobable – and we are back to Maxwell-
Boltzmann statistics.

22.6 Particle Trajectories and the Gibbs Paradox

There is more to learn from the failure of statistical independence apparent in the
Gibbs paradox. The Gibbs paradox is properly speaking a cluster of puzzles, where
the one so-far considered made direct contact with Einstein’s idea of statistical
independence. Here is another formulation.

24 Howard (1990), Norton (2006). Howard – and Norton – may well be right that the two were
connected in Einstein’s thinking; whether Einstein was right to connect them is the point I am
contesting.
25 For triviality, see Ghirardi et al. (2002), Ghirardi and Marinatto (2004), Penrose (2004); for

quantum trajectories, see Saunders (2006, p. 199–200, 2016, p. 169–70, 2018, p. 9–10); for Bell
inequalities, see Caulton (2020).
498 S. Saunders

If there is to be no entropy of mixing in the Gibbs set-up, there can be no new

(or comparatively few) microstates available after mixing, when the partition is
removed, than were not already present before. Yet it seems there are such states
available. Thus Versteegh and Dieks have recently argued26 :
Consider again two gas-filled chambers, each containing n identical particles. Before the
partition is removed the number of available states per particle is 12 V . After the partition
has been removed, the number of available states is V.. The reason is that after the partition’s
removal it has become possible for the particles to move to the other chamber. The
doubling of the number of available microstates thus expresses a physical freedom that was
not present before the partition was taken away. Trajectories in space-time have become
possible from the particles’ initial states to states in the other chamber. Particles from the
left and right sides can physically exchange their states.

Their conclusion is that according to classical theory an entropy of mixing is

unavoidable, and that an extensive entropy function, in classical theory, is no more
than a useful fiction. More recently, Dieks has argued indistinguishability in the
classical case amounts to invariance under particle relabelling, with no physical
consequences27 :
For example, two widely separated electrons, following distinct trajectories and labelled a
and b, still represent exactly the same physical situation after we have exchanged labels
(in the sense that the electron at the first position is now called electron b, and vice versa).
But this example also immediately demonstrates the physical insignificance of this type of
indistinguishability: in spite of the exchangeability of the labels, the particles themselves
can clearly be distinguished physically—we could easily tell them apart by a detection
device that is sensitive to positions. So it seems that this permutability argument, and
division by n! on its basis, only relates to our mode of description and does not possess
physical import. This division by n! cannot possibly lead to different predictions; which
makes it highly implausible that it could be relevant for a solution of the Gibbs paradox.

It appears that for indistinguishability of particles to have physical import it must

be impossible to tell them apart: it is indistinguishability in the sense of Boltzmann,
Planck, and Natanson, not Gibbs, Bose, and Einstein. But Dieks is right that division
by n! relates to our ‘mode of description’, specifically, to the choice of state space,
∼
whether, in the classical case, it is the reduced state space Γ or the specific state
space , and in the quantum case, whether it is the symmetrised subspace or the
full tensor product of Hilbert spaces. In all cases, with the entropy given by the
volume or dimensionality of the state space, it is highly plausible that this choice
will be relevant to Gibbs paradox. For consider: is it true, as Versteegh and Dieks
claim, that the number of available microstates doubles on removal of the partition?
Figure 22.1 is a space-time diagram for the Gibbs set-up. In Fig. 22.1b, the partition
remains in place throughout; in Fig. 22.1b, at time t , it is removed. New trajectories
are then possible after t . But does it follow that new states are made available at t2 ?
We may simplify further. Figure 22.2 is a spacetime diagram for the 2-particle
Gibbs paradox. Figure 22.2a shows an illustrative pair of trajectories when the

26 Versteegh and Dieks (2011, p. 742) (I have made minor changes of notation).
27 Dieks (2014, p. 1308).
22 What Price Statistical Independence? How Einstein Missed the Photon 499

a b

t t

t2 t2

t’

t1 t1
A B x A B x

Fig. 22.1 Space-time diagram for the mixing of gases. In (a) the partition remains in place, whilst
in (b) it is removed at time t

a b

Fig. 22.2 The 2-particle Gibbs paradox

partition is not removed, and Fig. 22.2b shows the kind of new trajectory possible
when it is, chosen to yield the same final positions and momenta. The intuition to
which Versteegh and Dieks appeal suggests that the two final microstates in Fig.
22.2a, b are different because they differ as to which particle has which position
and momentum. This is, fairly obviously, consistent with ordinary notions of
macroscopic objects, for macroscopic objects differ in their intrinsic, position- and
momentum-independent properties; but for microscopic particles, whose intrinsic
properties are exactly the same?
To insist that there is a difference even in this case is to insist on the use of the
specific phase, in Gibbs sense. In the space of specific phases of the two particles, a
continuous curve connecting two points related by particle interchange is a physical
process by which each particle acquires the position and velocity of the other. In the
space of specific states, it is an open curve whose end points are distinct. In the space
500 S. Saunders

of generic states, the initial and end points are one and the same. But the continuous
curve connecting them is not obliterated – the exchange has still taken place – it
is merely that it is now a closed curve in (reduced) phase space, with beginning
and end points numerically the same. (Of course these states obtain at different
times; as time slices they are numerically distinct; as such each may have a different
past, but that is accounted for by a change in the Hamiltonian, as time-dependent,
not the state.) On using the generic phase, it remains true that new trajectories are
possible, on removing the partition, but not new states; it is only that the same states
(microstates) can be approached in new ways.28
The Gibbs paradox, if it is a paradox, now reduces to this: are the final states of
Fig. 22.2a, b the same? Can the state of two particles be unchanged, when each of
the two particles has changed? No logical contradiction is involved. Imagine two
things, each of a definite shape, say that one is round and one is square; we can
imagine the square slowly turning into a circle, and the circle into a square, in such
a way that each occupies the same position as the other at the end.
It points, metaphysically, to the idea that the right notion of how the instantaneous
state of a joint system depends on the instantaneous state of its subsystems is global.
The state of the joint system supervenes on both – it is not built up from the state
of each subsystem considered individually. It may be this idea is more at home in
quantum theory, and perhaps there it is forced; but it may be the right metaphysics
more generally.
To see that it is forced in (conventional) quantum mechanics, consider a totally
symmetrised state of two identical particles, of the form:

1   

| Ψ  = √ |ϕ ⊗ | ψ + |ψ ⊗  ϕ
2

where |ϕ ∈ H is a one-particle state (and similarly | ψ, where | ϕand | ψ are
orthogonal). Let the particles be non-interacting, and suppose the unitary evolution
smoothly switches the two:

Û t |ϕ = |ψ ; Û t |ψ = | ϕ . (22.21)

Then the final state is unchanged:

Û t ⊗ Û t | Ψ  = | Ψ  .

Each of | ϕ and | ψ is (continuously) changed, yet the initial and final state is
the same.
The same question can be posed in terms of the Feynman sum-over-paths
formalism, using the same Fig. 22.2. In computing the transition amplitude between

28 Discounting the new states in which particle numbers on each side are changed, as already
mentioned.
22 What Price Statistical Independence? How Einstein Missed the Photon 501

initial and final states: do we sum the amplitudes for 2a and 2b coherently? We
should if the final state in the two cases is the same.
It was the same question posed, and answered, by Dirac in 1926, in terms of
the matrix mechanics, supplemented with Schrödinger’s new concept of state. He
concluded
If the positions of two of the electrons are interchanged, the new state of the atom
is physically indistinguishable from the original one. In such a case one would expect
only symmetrical functions of the coordinates of all the electrons to be capable of being
represented by matrices. It is found that this allows one to obtain two solutions of the
problem satisfying all the necessary conditions, and the theory is incapable of deciding
which is the correct one. One of the solutions leads to Pauli’s principle that not more than
one electron can be in any given orbit, and other, when applied to the analogous problem of
the ideal gas, leads to the Einstein-Bose statistical mechanics.29

Dirac explicitly considered the interchange of electrons in terms of a physical

process connecting the original state and the interchanged state. The two are the
same using Gibbs’ concept of generic phase, just as they are the same using Dirac’s
concept of symmetrised states; and indeed, only symmetric functions of the specific
phases can be represented by functions on the space of generic phases.
There is no question that in standard quantum mechanics, for identical particles,
initial and final states are identified, even if the process connecting them is a switch
of the form Eq. (22.21). If there is a rationally compelling argument to the contrary,
it awaits to be stated. But it may have seemed the onus was on the other side to
Einstein, Ehrenfest, and Lorentz, in that fateful first quarter of the last century. In
my opening remarks I suggested that there may have been a genuine misconception
at issue in their rejection of Gibbs’ concept of generic phase. Perhaps it was this:
they thought it meant, inter alia, that whether or not an interchange of like particles
had actually happened was somehow a meaningless question, according to Gibbs.
As realists, and given the existence of particle trajectories, they thought (rightly)
that that made no sense.
The view that quantum particle indistinguishability involves the renunciation
of particle trajectories has this in its favour. Departures from classical statistics
involves the absence of particle trajectories, understood as well-defined orbits of
1-particle states, for those departures involve more than one particle being found in
the same 1-particle state – prohibited if each has a well-defined and non-overlapping
trajectory.
But in another respect the verdict is quite different. The attribution of histories
to indistinguishable particles – or more precisely, the attribution of histories, or
changes of states, to light quanta, indeed, to an individual light quantum – the idea
of a light quantum as having a state-space at all – was Bose’s essential innovation
in 1924. It was this that broke the conceptual deadlock over the interpretation
of thermal radiation in terms of light quanta for every range of frequency and

29 Dirac (1926, p. 662). These are the two possibilities already remarked on in fn.11. Note that
parastatistics do not arise, unless we admit microstates that differ by particle permutation alone.
502 S. Saunders

temperature, and not just in the Wien regime. Had light quanta already been
recognised as non-independent in the Wien limit, the Rayleigh-Jeans regime might
better have been recognised for what it was: the regime in which the discrete
structure to radiation was no longer point-like, rather than non-independent; but
still movable.

References

Bach, A. (1997). Classical indistinguishable particles. Berlin: Springer.

Boltzmann, L. (1877). On the relationship between the second fundamental theorem of the
mechanical theory of heat and probability calculations regarding the conditions for thermal
equilibrium (K. Sharp & F. Matschinky, Trans.). Entropy 17 (2015) pp. 1971–2009. https://
doi.org/10.3390/e17041971.
Caulton, A. (2020). Entanglement by (anti-)symmetrisation does not Violate Bell’s Inequalities,
forthcoming.
Darrigol, O. (1986). The origin of quantized matter waves. Historical Studies in the Physical and
Biological Sciences, 16, 197–253.
Darrigol, O. (1991). Statistics and combinatorics in early quantum theory, II: Early symptoma of
indistinguishability and holism. Historical Studies in the Physical and Biological Sciences, 21,
237–298.
Dieks, D. (2014). The logic of identity. Distinguishability and indistinguishability in classical and
quantum physics. Foundations of Physics, 44, 1302–1316.
Dirac, P. (1926). On the theory of quantum mechanics. Proceedings of the Royal Society of London,
A112, 661–677.
Dorling, J. (1971). Einstein’s introduction of photons: Argument by analogy or deduction from the
phenomena? British Journal for the Philosophy of Science, 22, 1–8.
Ehrenfest, P., & Ehrenfest T. (1912). The conceptual foundations of the statistical approach in
mechanics (M. J. Moravcsik, Trans.). New York: Dover, 1990.
Ehrenfest, P., & Kamerlingh-Onnes, H. (1914). Simplified deduction of the formula from the theory
of combinations which Planck uses as the basis of his radiation theory. Proceedings of the
Amsterdam Acadamie, 17, 870–872. Reprinted in Paul Ehrenfest, Collected Scientific Papers,
Klein, M. (ed).; North-Holland, 1959, pp.353–56.
Einstein, A. (1905). On a heuristic point of view concerning the production and transformation of
light. In J. Stachel, D. Cassidy, J. Renn, & R. Schulmann (Eds.), [1990]: The Collected Papers
of Albert Einstein, Vol.2. The Swiss Years: Writings, 1900–1909. English Translation: Trans.
A. Beck, Princeton: Princeton University Press (1989).
Einstein, A. (1925). Quantum theory of the ideal monatomic gas. Second part. In D. K. Buchwald,
J. Illy, Z. Rosenkranz, T. Sauer, & O. Moses (Eds.), [2015]: The collected papers of Albert
Einstein, Vol. 14. English Translation. Princeton: Princeton University Press.
Epstein, P. (1936a). Gibbs’ methods in quantum statistics. In A. Haas (Ed.), A commentary on the
scientific writings of J. Willard Gibbs Vol. I: Theoretical physics (pp. 521–584). New Haven:
Yale University Press.
Epstein, P. (1936b). Critical appreciation of Gibbs’ statistical mechanics. In A. Haas (Ed.), A
commentary on the scientific writings of J. Willard Gibbs Vol. I: Theoretical physics (pp. 461–
518). New Haven: Yale University Press.
Ghirardi, G., & Marinatto, L. (2004). General criterion for the entanglement of two indistinguish-
able states. Physical Review A, 70, 012109.
Ghirardi, G., Marinatto, L., & Weber, Y. (2002). Entanglement and properties of composite
quantum systems: A conceptual and mathematical analysis. Journal of Statistical Physics, 108,
112.
22 What Price Statistical Independence? How Einstein Missed the Photon 503

Gibbs, J. W. (1902). Elementary principles in statistical mechanics. New York: Scribner.

Howard, D. (1990). “Nicht sein kann was nicht sein darf” or the prehistory of EPR, 1909–1935:
Einstein’s early worries about the quantum mechanics of composite systems. In A. I. Miller
(Ed.), Sixty-two years of uncertainty. New York: Plenum.
Mehra, J., & Reckenberg, H. (1987). The historical development of quantum theory: Vol. 5. New
York: Springer.
Norton, J. D. (2006). Atoms, entropy, quanta: Einstein’s miraculous argument of 1905. Studies in
History and Philosophy of Modern Physics, 37, 71–100.
Pais, A. (1982). Subtle is the Lord: The science and the life of Albert Einstein. Oxford: Oxford
University Press.
Penrose, R. (2004). The Road to Reality. London: Jonathan Cape.
Saunders, S. (2006). On the explanation of quantum statistics. Studies in the History and
Philosophy of Modern Physics, 37, 192–111. Available online at arXiv:quant-ph/0511136.
Saunders, S. (2013). Indistinguishability. In R. Batterman (Ed.), The Oxford handbook of philoso-
phy of physics (pp. 340–380). Oxford: Oxford University Press. http://philsci-archive.pitt.edu/
12448/.
Saunders, S. (2016). On the emergence of individuals in physics. In A. Guay & T. Pradeu (Eds.),
Individuals across the sciences. New York: Oxford University Press.
Saunders, S. (2018). The Gibbs paradox. Entropy, 20(8), 552. https://doi.org/10.3390/e20080552.
Saunders, S. (2020). The concept ‘indistinguishable’, forthcoming. Studies in the History and
Philosophy of Modern Physics.
Schrödinger, E. (1925). Bemerkungen über die statistische Entropiedefinition beim idealen Gas.
Preussische Akademie der Wissenschaften (Berlin), Proceeding (pp. 434–441).
Versteegh, M., & Dieks, D. (2011). The Gibbs paradox and the distinguishability of identical
particles. American Journal of Physics, 79, 741–746.
Chapter 23
How (Maximally) Contextual Is
Quantum Mechanics?

Andrew W. Simmons

Abstract Proofs of Bell-Kochen-Specker contextuality demonstrate that there

exists sets of projectors that cannot each be assigned either 0 or 1 such that each
basis formed from them contains exactly one 1-assigned projector. Instead, at least
some of the projectors must have a valuation that depends on the context in which
they are measured. This motivates the question of how many of the projectors must
have contextual valuations. In this paper, we demonstrate a bound on what fraction
of rank-1 projective measurements on a quantum system must be considered to
have context-dependent valuations as a function of the quantum dimension, and
show that quantum mechanics is not as contextual, by this metric, as other possible
physical theories. The figure of merit is a property only of the orthogonality graph,
or event structure, of a quantum scenario, which meshes with Pitowskys suggestion
that all relevant quantum behaviour can be captured by finite orthogonality graphs,
and raises questions about the same link Pitowsky explored between contextuality
and restrictions on quantum probabilities in the manner of Gleason’s theorem.

Keywords Quantum foundations · Quantum logic · Quantum contextuality ·

Maximal contextuality · Kochen-Specker contextuality

23.1 Introduction

The Bell-Kochen-Specker
  theorem tells us that there exist sets of projectors onto
states {ψ i }, such that it is impossible to assign to each of them a valuation
of 0 or 1 in a context-independent way, and have there be exactly one 1-valued
projector in every PVM context. We might ask the following question: under
an assumption of outcome-definiteness, how many such projectors can we give

A. W. Simmons ()
Department of Physics, Imperial College London, London, UK
e-mail: andrew@simmons.co.uk

© Springer Nature Switzerland AG 2020 505

M. Hemmo, O. Shenker (eds.), Quantum, Probability, Logic,
Jerusalem Studies in Philosophy and History of Science,
https://doi.org/10.1007/978-3-030-34316-3_23
506 A. W. Simmons

a noncontextual valuation to, and how many projectors (or more robustly, what
fraction of projectors), must be considered to have context-dependent valuations?
Historically, Kochen-Specker style proofs of contextuality, also known as strong
or maximal contextuality, have tended to have a “lynch pin” quality; removal of
any part of their structure causes the proof to fall apart in its entirety. In fact, this
fact has been exploited in order to tame the maximal contextuality demonstrated
by the Peres-Mermin magic square for all qubit states Bermejo-Vega et al. (2016),
as we need only remove the ability to measure one context before it fails to be
a proof of contextuality at all. For such proofs, then, since removing one context
is sufficient to remove all contextuality, certainly allowing a projector to vary in
its value assignment by context (by assigning it a C) will. This minimalism is
partially motivated by the fact that since the original paper by Kochen and Specker,
there has been interest in trying to find examples of small Kochen-Specker sets
(Arends et al. 2011). When only a single projector need be considered to take
a contextual valuation, finding a small such set is the only way of attaining a
higher fraction of contextual projectors. We will be looking  , for an assignment
to all projection operators onto a Kochen-Specker set {ψ i ψ i } → {0, C, 1}
such
 ,that the following rules hold: in any basis formed from the projectors onto
{ψ i ψ i }, there must be at least one with a C- or 1-valuation, and no more than
one with a 1-valuation. The fraction which must be given a C-valuation is the key
figure-of-merit that will be explored in this paper. This figure of merit, then, can
be seen as a generalisation or successor to the search for small Kochen-Specker
sets. Pitowsky (1998, 2005) explored the idea that Gleason’s theorem, viewed as a
proof of the Bell-Kochen-Specker theorem, can be thought of as providing more
quantitative, rather than qualitative, view on the Kochen-Specker theorem when
applied to quantum subtheories. Given any set of measurements, it is possible to
calculate the maximum deviation from the Born rule that the scenario permits;
thinking of a noncontextual 0,1-valuation as representing the maximum possible
deviation from the continuous Born rule probabilities. From this perspective, the
problem approached in this paper can be viewed as a complementary approach
to that of Pitowsky; we find our quantitative figure of merit by seeing how much
we have to relax the noncontextuality assumption to permit a classical model for a
certain compatibility structure, rather than how much the event structure can tolerate
deviations from quantum-like probabilities given by the Born rule.
In this paper, following Pitowsky’s approach (Pitowsky 1998), we will use
graph-theoretic methods as a computational tool to investigate how robust, in this
sense, quantum-mechanically accessible Kochen-Specker proofs can be. We will
demonstrate that it is possible that other physical theories can be more contextual,
given this definition, than quantum mechanics. We can consider this as a stronger
analogue than the observation that a PR-box is more nonlocal with respect to the
CHSH inequality than any quantum-mechanically accessible scenario. The figure
of merit in this paper can be defined referring only to this orthogonality graph,
which agrees with Pitowsky’s suggestion that all quantum behaviour should be
viewed as a product of the compatibility structure of quantum measurements, via the
restrictions imposed on possible probability rules by Gleason’s theorem. The central
23 How (Maximally) Contextual Is Quantum Mechanics? 507

question we ask and partially resolve in this paper explores the limits of Pitowsky’s
ideas and work on logical indeterminacy (Pitowsky 1998); unlike in the situation
Pitowsky considered, we see that allowing some observables to have valuations that
contextually vary allows many others to be deterministic. We are then able to query
to what extent the logical indeterminacy principle remains true, even in the presence
of explicit contextuality.
One metric for contextuality is the contextual fraction Abramsky et al. (2017),
in which we consider our probability distribution to be a convex mixture between
a noncontextual theory and a contextual one. This notion is what prompts the
term “maximal contextuality”; if a scenario is maximally contextual then it has
a contextual fraction that is the highest possible, viz 1. It is independent under
a re-labelling of the measurement effects, but the contextual fraction is hard to
calculate when the number of corners of the noncontextual polytope is large, which
is true in many natural situations such as in nonlocality scenarios in which one
party has access to measurements with three or more outcomes Simmons (2018).
The figure-of-merit explored in this paper acts as a complementary measure of
contextuality that can distinguish different scenarios which have the maximum
possible contextual fraction of 1. However, it is likely to be hard to calculate in many
circumstances, and in the case in which one is using the entirety of Hilbert space as
one’s vector set, we shall explore the connection to open problems in mathematics
known as Witsenhausen’s problem (Witsenhausen 1974) and the Hadwiger-Nelson
problem (Soifer 2008).

23.2 How Robust Can Quantum Maximal Contextuality Be?

Consider the orthogonality graph G of a maximal nonlocality scenario, inwhich   we

identify projectors with vertices, and two vertices g , g , identified with ψ 1 , ψ 2
,   1 2
are joined by an edge iff ψ 1 ψ 2 = 0. Following Renner and Wolf Renner and
Wolf (2004), we will restrict our attention to outcome assignments in which we do
not ever assign a “1” valuation to two orthogonal vectors; this is known as a weak
Kochen-Specker set. A weak Kochen-Specker
  set can be extended to a full Kochen-
Specker set with the addition of O(|ψ i |2 d) extra vectors, or we can consider the
incomplete bases as being completed by additional projectors, whose valuations are
allowed to be contextual as they are outside the scope of our scenario. In fact, this
distinction will not be crucial for any of the main results; it does however make
certain definitions conceptually clearer. An independent set is a subset of vertices,
none of which are joined by an edge. Any noncontextual assignment of “1”s, then,
will form an independent set since we cannot have two orthogonal vectors, (which
correspond to two vertices joined by an edge) in our set of vectors which are
noncontextually given “1” assignments. The size of the largest independent set of a
graph G is denoted α(G) and is called the independence number. Another important
concept will be that of a clique:
508 A. W. Simmons

Definition 1 (Clique) For a graph G = {V , E}, vertex subset C ⊂ V is a Clique if

the subgraph induced from G on C is complete, that is, every vertex in C is connected
to every other; and C is inclusion-maximal.
We say that a clique is a maximum clique if there is no clique in G with more
vertices. The size of a maximum clique in G is denoted ω(G), and so we note that
for a d-dimensional quantum system, as long as the set of vectors we can measure
includes at least one basis, we will have ω(G) = d. In a hidden variable model
for a scenario, we require that in every context, there be at least one vector that can
take a “1” value either contextually or noncontextually. This motivates the following
definition:
Definition 2 (Maximum-clique hitting-set) For a graph G = {V , E}, a vertex
subset T ⊂ V is a maximum-clique hitting-set or maximum-clique transversal if
for all cliques C, |C| = ω(G), T ∩ C = ∅.
In any contextual hidden-variable model, the set of vertices associated with a set
of vectors given either noncontextual or contextual “1” assignments is a Maximum-
clique hitting set, since every context needs an outcome. This allows us to define
our first figure-of-merit for a set of vectors forming a Kochen-Specker type proof of
the Bell-Kochen-Specker theorem.

qs (G) = min min |T | − |A|, (23.1)

T A⊂T

where T is a maximum-clique hitting-set. We see that this is a graph-theoretic

representation of the figure-of-merit we posited above: we take a set T of vectors
which will recieve a “1” valuation, either contextually or noncontextually, and
subtract from it the number of those vectors that can support a noncontextual
valuation. This quantity then represents the number of vectors that must receive
a contextual valuation in any model for the scenario. However, we can note that it is
possible to achieve an arbitrarily high value of qs just by taking an equal number of
noninteracting copies of, for example, the Peres-Mermin magic square. In order that
we might more directly compare this quantity between graphs of different sizes, we
can consider its graph density:

qs (G)
q(G) = . (23.2)
|V (G)|

We now have a graph-theoretic formulation of our quantity of interest. What values

of this quantity are obtainable, either by quantum-mechanically realisable scenarios,
or in general probabilistic theories? We note the following trivial lower bound:

qs (G) ≥ min |T | − α(G). (23.3)

We note that in general, it is NP-complete to calculate the value of min |T | if we

know ω(G), since for triangle-free graphs, this is equivalent to finding a vertex cover,
23 How (Maximally) Contextual Is Quantum Mechanics? 509

Fig. 23.1 Three images illustrating the calculation of the figure of merit. The first is the orthogo-
nality graph for the 33-ray proof of the Kochen-Specker theorem by Peres (1991). Highlighted
in the second in red is a maximal independent set, that represents a maximal assignment of
noncontextual 1-valuations. In fact, only a single context lacks a 1-valued vector, visible in the
centre of the graph. Highlighted in the third in blue is a projector that can be given a contextual
valuation in order that every context contain a ray assigned either a 1 or a contextual valuation; it
will take a 1 valuation in the “central” context, and a 0 in all others

which is NP-complete (Poljak 1974). Also, it is NP-complete to calculate α(G)

(Garey and Johnson 1979). This suggests that calculation of this quantity may be
computationally difficult. A tension exists in trying to optimise q(G). Particularly
well-connected graphs may achieve an exponentially small independence density –
this is true for a few of the families of graphs considered in this paper – so few
vectors can receive a noncontextual assignment of 1, but this same connectedness
means that allowing a single vector’s valuation to vary contextually “fixes” many
contexts (Fig. 23.1).
We will now demonstrate an upper bound on the quantity q(G) for any quantum-
mechanically accessible G in terms of the quantum dimension of the vectors that
give rise to it.
 
Theorem 1 Any orthogonality graph G induced by a set of vectors ψ i in a d-
dimensional Hilbert space has
 
1 d−1 1
q(G) ≤ 1 − − d−1 . (23.4)
d 2

Proof The proof proceeds  by

 demonstrating an explicit procedure for finding a T
and an A ⊂ T , for any {ψ i }, with upper-bounded worst-case performance.
The set of vectors |φ ∈ Hd such that |ψ|φ|2 ≥ t for some t ∈ R forms a closed
complex hyperspherical cap centred at |ψ. Consider a basis for Hd . Without loss
of generality, we can take this to be the standard basis for the set. By symmetry
 the quantity mini |ψ|i| is maximised (albeit nonuniquely) by |ψ =
concerns,
d −1/2 i |i. This puts an upper bound on the size of the largest hyperspherical
angle that can separate an arbitrary vector from its closest
 basis element.
 A closed
complex hyperspherical cap subtending this angle, |ψ|ψ|φ|2 ≥ 1/d , will then
510 A. W. Simmons

necessarily capture at least one element from the basis, and so any set of vectors
within such a hyperspherical cap C construes a maximum-clique hitting-set, TC .
Since columns of a Haar-random unitary matrix over Cd are uniformly random
unit vectors, the volume V of this hyperspherical cap as a proportion of that of the
sphere is given by 1 − F (t), where F (t) is the cumulative distribution function of
the squared modulus of an entry in such a Haar-random unitary. This quantity is
well known in the field of random matrix theory and it can be found, for example,
as a special case of a result by Życzkowski and Sommers (2000), and will be
demonstrated in Sect. 23.3.
V
= (1 − t)d−1 . (23.5)
VS

Clearly, the value of q(G) for a quantum graph is upper bounded by the size of any
maximum-clique hitting set, so if we can bound the minimum number of vectors
we must intersect with such a hyperspherical cap, this also constitutes a bound on
the value q(G). However, we would not expect this bound to be tight, since we
have ignored the independent set subtrahend entirely. We would like to demonstrate
an independent set within our maximum-clique hitting-set TC , and therefore give
a better upper bound on the quantity qs (G) = minT minA⊂T |T | − |A|, giving an
explicit construction procedure for a T and A ⊂ T .
We note that a hyperspherical cap containing the vectors |φ such that |ψ|φ|2 >
1/2 cannot contain two orthogonal vectors as a simple application of the triangle
inequality. We can place this smaller hyperspherical cap anywhere within the larger
hyperspherical cap, and this will form an independent set A that is a subset of TC .
For the rest of this proof, we will assume that the two hyperspherical caps are centred
on the same point; this will make no difference to the worst case scenario which
forms the bound. The region of Hilbert space, then, in which vectors count towards
the figure of merit, is a member of a two-parameter family of annuli on the complex
hypersphere, illustrated in Fig. 23.2 and defined by

t1 ≤ |ψ|φ|2 ≤ t2 . (23.6)

Fig. 23.2 An illustration of 1

how this construction splits
up the Hilbert space into
different regions, that can be C
associated with projectors
given noncontextual
1-assignments, contextual
projectors, and projectors
given noncontextual
0
0-assignments
23 How (Maximally) Contextual Is Quantum Mechanics? 511

Such an annulus takes up a proportion

p(t1 , t2 ) = (1 − t1 )d−1 − (1 − t2 )d−1 (23.7)

of the Hilbert space. Our construction has t1 = 1/d , and t2 = 1/2, giving us a
hyperspherical annulus defined by 1/d ≤ |ψ|φ|2 ≤ 1/2. It now remains to be
proved that no matter the positions of the set of vectors {ψ}, there must be some
place in which we can place this annulus such that we only intersect strictly less
than p(t1 , t2 ) of them.
Consider the canonical action of SU (d) on the unit hypersphere in Cd , which we
shall denote SCd−1 . We define a space SU (d) × SCd−1 , where SU (d) is equipped with
the Haar measure, and SCd−1 is equipped with the uniform measure. Equivalently,
we could have taken Hd rather than SCd−1 , in which case it would be equipped with
the Fubini-Study
  metric.
At each ψ i , we place a bump function f|ψ  with unit integral and support only
  i
within a disc of radius  around ψ i . Let At ,t denote the hyperspherical annulus
1 2
{|φ | t1 ≤ |0|φ|2 ≤ t2 }. By Fubini’s theorem, we have:
     
1g.z∈At1 ,t2 f|ψ  (z) dz dg= 1g.z∈At1 ,t2 f|ψ  (z) dg dz
d−1 i d−1 i
SU (d) SC i SC SU (d) i
(23.8)
We note that if we extend our annulus’s extent by , so that we have t1 →
t1 −  and t2 → t2 + , we capture the entirety of the measure of the bump
functions
  centred
 insidethe -original annulus. Hence,
 ψ i ψ i ∈ g −1 At ,t  ≤ d−1 1g.z∈A  for all g ∈ SU (d) we have
1 2 S C
t1 −,t2 + i f|ψ  (z) dz, and so we have
i

 
      
   
 ψ i  ψ i ∈ g −1 At1 ,t2  dg ≤ 1g.z∈At1 −,t2 + f|ψ  (z) dg dz,
d−1 i
SU (d) SC SU (d) i
(23.9)
Rearranging:
 
     
   
 ψ i  ψ i ∈ g −1 At1 ,t2  dg ≤ f  (z) 1g.z∈At1 −,t2 + dg dz
d−1 |ψ i 
SU (d) i SC SU (d)
(23.10)
Twirling the indicator function 1g.z∈At1 −,t2 + with respect to the Haar measure
over SU (n) reduces it to a constant function with value p(t1 − , t2 + ), the
proportion of the Hilbert space taken up by the annulus.
 
    
   −1 
 ψ i  ψ i ∈ g At1 ,t2  dg ≤ p(t1 − , t2 + ) f|ψ  (z) dz
d−1 i
SU (d) i SC
(23.11)
 
≤ p(t1 − , t2 + ) ψ i  (23.12)
512 A. W. Simmons

Since the left hand side of Equation 23.11 represents  an average
 over the rotation
group, there must be some g ∈ SU (d) such that  ψ i ψ i ∈ g −1 At ,t  ≤ p(t1 −
  1 2
, t2 + ) ψ i , and this is true for all . Therefore:

q(G) ≤ lim p(t1 − , t2 + ) (23.13)

→0
 
1 d−1 1
≤ 1− − d−1 . (23.14)
d 2

*
)
Corollary 1 Any quantum mechanically accessible G has q(G) ≤
4251920575/11019960576∼ 0.385838.
 d−1
Proof Calculating the derivative of 1 − d1 − 2d−1
1
with respect to d shows
that it takes a maximum between d = 9 and d = 10. Evaluation of the quantity
at these points reveal the maximum to be at d = 9, when we get (1 − 1/9)8 −
1/28 = 4251920575/11019960576. This, then, forms a hard limit of q(G) for any quantum
mechanically achievable G. *
)
We note that as the quantum dimension approaches infinity, we achieve a limiting
value of 1/e, so quantum systems with very high dimension are viable candidates
for providing robust contextuality scenarios. In Fig. 23.3, we see the bound on q(G)
plotted as a function of d.
Above, it was shown that a spherical cap taking up a proportion of (1 − 1/d )d−1
of Hilbert space must capture at least one vector from every orthonormal basis.
However, if for many bases this spherical cap captures more than one basis element,
this might suggest that this spherical cap was not the most efficient way of choosing
a vector from each basis. We note, though, that the exact geometry of the set

• • • • • • • • • • • •
• • •
1
e •
0.3 •

0.2 •

0.1

• d
5 10 15 20

Fig. 23.3 A chart showing the value of the bound on q(G ) for quantum-mechanically accessible
G as a function of the Hilbert space dimension d
23 How (Maximally) Contextual Is Quantum Mechanics? 513

of projectors will have a large effect on the efficiency of this construction. One
one hand, an increase in the quantum dimension allows for more complicated
structures and interplay of the permitted projectors that, as we shall demonstrate,
seems to allow for construction of quantum graphs with very small independence
number. However, it also has a dampening affect on the quantity |T | since the bases
themselves have more geometrical constraints.
In fact such an annulus can in the worst case can contain every vector from
a basis. Below, we will see a specific example of a set of vectors providing a
nonlocality scenario, in which every projector is captured by a pessimally-chosen
spherical cap position. In d-dimensional Hilbert space, a spherical cap of proportion
(1− 1/n)d−1 can capture n vectors out of a basis. So for any chosen n , as d increases,
we see that the fraction of the spherical cap in the construction that cannot capture
n of a basis tends to 0. However, if we consider the fraction of the spherical cap
that can capture a constant proportion of basis elements c, we get the following
behaviour under increasing d:
 
1 d−1 1
lim 1 − = e− c . (23.15)
d→∞ cd

This might imply that the quality of the bound does not degrade as the dimension
increases (Fig. 23.4).

Quantum Transversal Independence Figureof

Name
dimension size number merit
√ √  
ABK triangle-free log n
graphs (Alon et al. 2010)
n − c n log n c n log n 1−O n

Quantum upper 
1 d−1
d ≤n 1− d ≥ n
2d−1
≤0.385838
bound
E8 root
8 ≤ 11n
60
n
15 ≤0.116̇
system
Two-qubit stabiliser 3n n
4 10 5 0.1
quantum mechanics
Peres-Mermin magic 7n 5n
square (Peres 1991)
4 24 24 0.083̇
Cabello’s 18-ray 5n 4n
4 18 18 0.05̇
proof (Cabello 1997)
Peres’s 33-ray 9n 12n
3 33 33 0.0̇3̇
proof (Peres 1991)

Fig. 23.4 A table comparing the graph constructions in this paper; each concrete example
represents the graph union of multiple noninteracting copies of the KS proof. Two-qubit stabiliser
quantum mechanics forms a KS proof using 60 rays and 105 contexts; the E8 root system forms a
KS proof using 120 rays and 2025 contexts. The quantities related to each were calculated using
a proof by exhaustion, or are the best found during a non-exhaustive search, such as is the case
for the E8 system. It is perhaps of note that the lower bound of the size of a traversal less the
independence number forms a trivial bound only for Peres’s 33-ray proof, and in fact is optimal
for Cabello’s 18-ray proof and the Peres-Mermin magic square
514 A. W. Simmons

23.2.1 Triangle-Free Graphs

In triangle-free graphs, the maximum cliques are merely edges, and so each context
consists of two measurement objects. Quantum mechanically, the only such graphs
are uninteresting: the disjoint graph union of K2 graphs. These do not have enough
structure to form a proof of the Bell-Kochen-Specker theorem; this reflects the fact
that each projector appears in only a single context. Many such graphs, then, are not
quantum-mechanically accessible. However, they do allow us to prove a concrete
bound on what values of q(G) are achievable within the framework of generalised
probabilistic theories. In such graphs, the concept of a maximum-clique hitting-set
reduces to that of a covering of edges by vertices, known as a vertex cover. The size
of the minimal vertex cover of G is denoted τ (G). This reduction of our figure-of-
merit allows us to make use of a bound derived for Ramsey type problems.
In the triangle-free case, we have a powerful relation between maximum-clique
hitting-sets and independent sets; namely that they are graph complements of each
other. A maximum-clique hitting-set, or vertex cover, is a set of vertices T such that
for every edge, at least one of that edge’s vertices is in T . We can see, then, that the
set V − T must be independent, since if there were two vertices in V − T connected
by an edge, then neither of those vertices would be in T , a contradiction. Conversely,
if we have an independent set A, then V − A must be a vertex cover; if there were
an edge with neither vertex in V − A, then both vertices are in A, a contradiction.
This means that we can bound qs (G) as

qs (G) ≥ min |T | − α(G) = |V (G)| − 2α(G). (23.16)

G

In other words, to be able to lower bound qs (G), we need only consider the
independence number, α(G). We seek, therefore, triangle-free graphs with low
independence number. We can invoke here a theorem due to Alon et al. (2010).
Theorem 2 There exists a constant c such that for all n√∈ N there exists a regular
triangle-free graph Gn , with V (Gn ) = n, and α(Gn ) ≤ c n log n.
Hence, we have
√
n − 2c n log n
lim q(Gn ) ≥ lim (23.17)
n→∞ n→∞ n

log n
≥ 1 − 2c lim = 1. (23.18)
n→∞ n

Therefore, q(Gn ) = 1 is approachable in the limit of n → ∞. We see then that

quantum mechanics does not display maximal contextuality as robustly, given this
definition of robustness, as other possible GPTs. This could potentially have an
impact on attempts to recreate quantum mechanics as a principle theory.
23 How (Maximally) Contextual Is Quantum Mechanics? 515

23.3 Higher-Rank PVMs

We now extend the result to apply to projective valued measures including projectors
of rank greater than 1, although we are still considering the case in which contexts
are made up of a set of projectors which sum to the identity, rather than the more
general case of a set of commuting projectors. We will allow a

rank-k.projector P to
 (i)
“inherit” a labelling as a function of the assignments to some ψ P , which form
  (i) ./ (i) 
a specially chosen decomposition of P as P = i ψ P ψ P .
Given a variable assignment to states f : Hd → {0, C, 1}, we can define a
variable assignment to all projection operators on Hd , g : P(Hd ) → {0, C, 1} by
the following:
⎧

(i)
   (i) ./ (i)  
 (i)
.
⎪
⎪ 0 if ∃ ψ P , P = i ψ P ψ P , f ψ P = 0 ∀i,
⎨
   (i) ./ (i)  

.
g(P ) = 1 elif ∃ ψ P , P = i ψ P ψ P , f ψ P
(i) (i)
∈ {0, 1} ∀i,
⎪
⎪
⎩
C o/w.
(23.19)
The motivation for this notion of value inheritance is as follows: in any context
containing P , we can consider replacing it with one of its maximally fine-grained
decompositions, which then inherits value assignments from f derived via the
annulus method. Since the assignment of values to this rank-1 decomposition of
P is necessarily consistent with the restrictions of the scenario, we can consider
a post-processing that consists of a coarse-graining of those fine-grained results,
“forgetting” which one of them occurred. This process must also be consistent.
Hence, as long as there exists some decomposition of P into rank-1 projectors, each
of which is assigned a noncontextual valuation, then P can inherit a noncontextual
valuation. However if there must be a contextual vector in any such decomposition,
we must treat P as having a contextual valuation.
The challenge, then, is to characterise which projectors P must be given a
contextual valuation under this schema, and then prove a bound analogous to that of
Theorem 1.
Theorem 3 In a scenario in which the available measurements are rank-r projec-
tors {Pi }, performed on a quantum system of dimension d, then at most (I1/2 (r, d −
r) − Ir/d (r, d − r))|{Pi }| are given a contextual valuation, where Ix (α, β) is the
regularised incomplete beta function.
Proof The proof proceeds similarly to the proof of Theorem 1. We will identify
a generalised annulus within the space of rank-r projectors on Hd which can be
associated with contextual valuations, and then use its volume in proportion to that
of the overall Hilbert space to form a bound on how many must be given contextual
valuations.
516 A. W. Simmons

Lemma 1 Under the extension of a valuation of rank-1 projectors given by the

annulus method from Theorem 1 centred on |φ, a rank-r projector Pr is contextual
only if r/d ≤ Tr(Pr |φφ|) ≤ 1/2.
Proof of Lemma First, consider the case that Tr(Pr |φφ|) > 1/2. We could
equivalently write this as φ|Pr |φ = φ|Pr Pr |φ > 1/2. We note that Pr |φ is
a vector in the eigenvalue-1 subspace of Pr . As such, we can find a basis for Pr
that includes Pr |φ, up to a normalisation constant. By design, each of these other
basis elements |i will have i|φ = i|Pr |φ = 0. Hence, this corresponds to a
decomposition of Pr in which every vector would be given either a 1 or a 0 valuation,
and so Pr inherits a 1 valuation.
Next, we consider the case that Tr(Pr |φφ|) = r/d −  < r/d . We wish to
decompose Pr in such a way that each of the elements |i of the decomposition
have |i|φ| < 1/d . Again, we consider the vector Pr |φ and choose a basis for the
1-eigenspace of Pr so that we have
& '

r−1
1 
Pr |φ = − |i. (23.20)
d r
i=0

Each of these elements in the decomposition would recieve a noncontextual 0

valuation, so Pr inherits a 0 valuation. This concludes the proof. )
*
Using the result of the lemma, we can apply a proof method identical to that in
Theorem 1 if we can calculate the proportion of the Hilbert space taken up by the
set {Pr |r/d ≤ Tr(Pr |φφ|) < 1/2}.
As before, entries of a uniformly random unit vector in Cd have the same
distribution as the entries of a column in a Haar-random d × d unitary matrix.
Applying a result of Życzkowski
 and Sommers (2000), if U is a Haar-random d × d
r  2
unitary, then defining Y = k=1 Uk,1  , we have

Y
Pd,r (y) = cd,r y 2r−1 (1 − y 2 )d−r−1 , (23.21)

Y (y) is the probability density function for the random variable Y , and
where Pd,r

2 2(d)
cd,r = = , (23.22)
B (r, d − r) (r)(d − r)

in which B is the Euler beta function. and  is the gamma function. Performing a
change of variables, then, in which T = Y 2 , we get:
  
 d √  Y √ 
T
Pd,r (t) =  t  Pd,r t (23.23)
dt
23 How (Maximally) Contextual Is Quantum Mechanics? 517

t r−1 (1 − t)d−r−1
= (23.24)
B(r, d − r)

This is exactly a Beta distribution with shape parameters α = r and β = d − r; we

have derived that T ∼ Beta(r, d − r). We note taking r = 1, and integrating, we
recover our earlier result used in Theorem 1. Since the CDF for a Beta distributed
random variable is given by the regularised incomplete beta function, Ix (α, β), the
proportion of the Hilbert space taken up by the set {Pr |r/d ≤ Tr(Pr |φφ|) ≤ 1/2} is
given by

I1/2 (r, d − r) − Ir/d (r, d − r). (23.25)

To complete the proof we once again apply the argument from Fubini’s theorem
used in the proof of Theorem 1. *
)
Corollary 2 For any choice of d, r, the proportion of projectors requiring a
contextual valuation is always less than 1/2.
Proof We have for any d, r, that the proportion of projectors requiring a contextual
valuation is bounded above by I1/2 (r, d − r) − Ir/d (r, d − r). We wish, then, to bound
this quantity above by 1/2.
Since Ix (α, β) is the cumulative distribution function for a random variable with
distribution Beta(α, β), a trivial upper bound for this quantity is given by

I1/2 (r, d − r) − Ir/d (r, d − r) < 1 − Ir/d (r, d − r). (23.26)

In fact for large d, this is a reasonable approximation, since we can apply

Chebyshev’s inequality to show that I1/2 (r, d − r) ≥ 1 − O(d −2 ). We wish, then, to
lower bound the value of Ir/d (r, d − r). The mean-median-mode inequality (Kerman
2011) for the Beta distribution with α ≤ β, which is met here since r must be a
nontrivial factor of d, is given by

α−1 α
≤ m(α, β) ≤ , (23.27)
α+β −2 α+β

where m(α, β) represents the median of a random variable with a Beta(α, β)

distribution. This is an equality only when α = β; although this corresponds to
a trivial contextuality scenario. Importantly, we have
r
m(r, d − r) ≤ . (23.28)
d

Hence by the monotonicity of Ix (α, β), we have Ir/d (r, d−r) ≥ Im(r,d−r) (r, d−r) =
1/2,
and the result follows.
*
)
518 A. W. Simmons

23.4 Conclusion and Future Work

There is a large gulf between the best-known realisation of a quantum-mechanically

accessible BKS proof and the bounds on the robustness of such a proof that we
have derived in this article. However, it may be that to prove a tighter bound on
what is quantum-mechanically possible in a way similar to that of this paper is
mathematically difficult; this is because the method requires that we find a set that
is independent, regardless of the specific structure of the set of vectors in question.
This implies bounds on the independence number of the associated orthogonality
graph for all finite contextuality scenarios, and by extension also the contextuality
scenario that includes every quantum state. This problem, to choose the largest set
of points on a hypersphere such that no two points in the set are orthogonal, is
essentially a problem known as Witsenhausen’s problem (Witsenhausen 1974) after
the author of the paper that formulated this problem, and proved the trivial bound of
1/d . It has been conjectured that the spherical cap approach is optimal for both the
real case, as conjectured by DeCorte and Pikhurko (2016) by Gil Kalai and in the
complex case by Montina. However, this is not even known for the original three-
real-dimensional case considered by Kochen and Specker. In turn, the solution to
Witsenhausen’s problem implies lower bounds for the Hadwiger-Nelson problem
(DeCorte and Pikhurko 2016): the number of colours needed to colour Euclidean
spaces if no two points distance 1 away from each other can have the same colour.
The solution to this problem is thought to depend on the specific model of ZF set
theory adopted (Soifer 2008). Additionally, if a better construction is found, then we
might also have to vary the shape of the area designed to capture at least one element
from each basis, so that this latter area is a strict superset of the former, meaning that
this efficiency may not be directly capturable by a variation of this proof. However,
a more efficient construction would lead to a better bound in the high-dimensional
limit.
In this paper, we considered contexts formed of projective matrices that sum to
the identity. Recently, attention has been drawn to an analogous problem defined
for sets of mutually commuting Hermitian matrices, and so an extension of variant
of this result in that case would be helpful for bounding the amount of classical
memory needed for an unbiased weak simulation of quantum subtheories (Karanjai
et al. 2018).
A comparison between this problem and that explored by Pitowsky reveals
potential connections that could be investigated, the most obvious of which is to
investigate the extent to which the two figures of merit capture the same information
about the contextuality of a quantum subtheory. They have similar properties, such
as being defined as a function of just the compatibility structure of a scenario given
by orthogonality graphs; and taking noncontextual scenarios as an extreme value.
Such a comparison would be restricted only to connected graphs since Pitowsky’s
figure of merit is not naturally invariant under the addition of a rotated version of a
scenario to itself, when the two copies have no commensurability.
23 How (Maximally) Contextual Is Quantum Mechanics? 519

Acknowledgements Special thanks to Mordecai Waegell, who suggested using the E8 root
system and two-qubit stabiliser quantum mechanics to demonstrate lower bounds on what values
of the figure of merit in Sect. 23.2 were quantum-mechanically accessible, as well as his help
calculating this lower bound. Special thanks also to Terry Rudolph for asking this question in
the first place. I am grateful to Angela Xu, Angela Karanjai and Baskaran Sripathmanathan for
their helpful discussions which have guided this project. I acknowledge support from EPSRC, via
the CDT in Controlled Quantum Dynamics at Imperial College London; and from Cambridge
Quantum Computing Limited. This research was supported in part by Perimeter Institute for
Theoretical Physics. Research at Perimeter Institute is supported by the Government of Canada
through the Department of Innovation, Science and Economic Development and by the Province
of Ontario through the Ministry of Research and Innovation.

References

Abramsky, S., Barbosa, R. S., & Mansfield, S. (2017). The contextual fraction as a measure of
contextuality. arXiv:1705.07918.
Alon, N., Ben-Shimon, S., & Krivelevich, M. (2010). A note on regular Ramsey graphs. Journal
of Graph Theory, 64(3), 244–249.
Arends, F., Ouaknine, J., & Wampler, C. W. (2011). On searching for small Kochen-Specker vector
systems (pp. 23–34). Berlin/Heidelberg: Springer.
Bermejo-Vega, J., Delfosse, N., Browne, D. E., Okay, C., & Raussendorf, R. (2016). Contextuality
as a resource for qubit quantum computation. arXiv:1610.08529.
Cabello, A. (1997). A proof with 18 vectors of the Bell-Kochen-Specker theorem (pp. 59–62).
Dordrecht: Springer.
DeCorte, E., & Pikhurko, O. (2016). Spherical sets avoiding a prescribed set of angles.
International Mathematics Research Notices, 2016(20), 6096–6117.
Garey, M. R., & Johnson, D. S. (1979). Computers and intractability; a guide to the theory of
NP-completeness. W.H. Freeman and Company. New York.
Karanjai, A., Wallman, J., & Bartlett, S. (2018). Contextuality bounds the efficiency of classical
simulation of quantum processes. arXiv:1802.07744.
Kerman, J. A closed-form approximation for the median of the beta distribution. arXiv:1111.0433,
November 2011.
Peres, A. (1991). Two simple proofs of the Kochen-Specker theorem. Journal of Physics A:
Mathematical and General, 24, 175–178.
Pitowsky, I. (1998). Infinite and finite Gleason’s theorems and the logic of indeterminacy. Journal
of Mathematical Physics, 39, 218.
Pitowsky, I. (2005). Quantum mechanics as a theory of probability. In W. Demopoulos &
I. Pitowsky (Eds.) Physical theory and its interpretation (The western Ontario series in
philosophy of science, Vol. 72). Dordrecht: Springer.
Poljak, S. (1974). A note on stable sets and colorings of graphs. Commentationes Mathematica
Universitatis Carolinae, 15(2), 307–309.
Renner, R., & Wolf, S. (2004). Quantum pseudo-telepathy and the Kochen-Specker theorem. In
International Symposium on Information Theory, 2004. ISIT 2004. Proceedings (pp. 322–322).
Simmons, A. W. (2018, February). On the computational complexity of detecting possibilistic
locality. Journal of Logic and Computation, 28(1), 203–217. https://doi.org/10.1093/logcom/
exx045.
Soifer, A. (2008). The mathematical coloring book: Mathematics of coloring and the colorful life
of its creators. New York/London: Springer.
Witsenhausen, H. S. (1974). Spherical sets without orthogonal point pairs. American Mathematical
Monthly, 81, 1101–1102.
Życzkowski, K., & Sommers, H.-J. (2000). Truncations of random unitary matrices. Journal of
Physics A: Mathematical and General, 33(10), 2045.
Chapter 24
Roots and (Re)sources of Value
(In)definiteness Versus Contextuality

Karl Svozil

Abstract In Itamar Pitowsky’s reading of the Gleason and the Kochen-Specker

theorems, in particular, his Logical Indeterminacy Principle, the emphasis is on the
value indefiniteness of observables which are not within the preparation context.
This is in stark contrast to the prevalent term contextuality used by many researchers
in informal, heuristic yet omni-realistic and potentially misleading ways. This paper
discusses both concepts and argues in favor of value indefiniteness in all but a con-
tinuum of contexts intertwining in the vector representing a single pure (prepared)
state. Even more restrictively, and inspired by operationalism but not justified by
Pitowsky’s Logical Indeterminacy Principle or similar, one could identify with a
“quantum state” a single quantum context – aka the respective maximal observable,
or, in terms of its spectral decomposition, the associated orthonormal basis – from
the continuum of intertwining context, as per the associated maximal observable
actually or implicitly prepared.

Keywords Logical indeterminacy principle · Contextuality · Conditions of

possible experience · Quantum clouds · Value indefiniteness · Partition logic

24.1 Introduction

An upfront caveat seems in order: The following is a rather subjective narrative of

my reading of Itamar Pitowsky’s thoughts about classical value indeterminacy on
quantum logical structures of observables, amalgamated with my current thinking
on related issues. I have never discussed these matters with Itamar Pitovsky
explicitly; therefore the term “my reading” should be taken rather literally; namely

K. Svozil ()
Institute for Theoretical Physics, Vienna University of Technology, Vienna, Austria
e-mail: svozil@tuwien.ac.at; http://tph.tuwien.ac.at/~svozil

© Springer Nature Switzerland AG 2020 521

M. Hemmo, O. Shenker (eds.), Quantum, Probability, Logic,
Jerusalem Studies in Philosophy and History of Science,
https://doi.org/10.1007/978-3-030-34316-3_24
522 K. Svozil

as taken from his publications. In what follows classical value indefiniteness on col-
lections of (intertwined) quantum observables will be considered a consequence, or
even a synonym, of what he called indeterminacy. Whether or not this identification
is justified is certainly negotiable; but in what follows this is taken for granted.
The term value indefiniteness has been stimulated by recursion theory (Rogers,
Jr. 1967; Odifreddi 1989; Smullyan 1993), and in particular by partial func-
tions (Kleene 1936) – indeed the notion of partiality has not diffused into physical
theory formation, and might even appear alien to the very notion of functional value
assignments – and yet it appears to be necessary (Abbott et al. 2012, 2014, 2015)
if one insists (somewhat superficially) on classical interpretations of quantized
systems.
Value indefiniteness/indeterminacy will be contrasted with some related inter-
pretations and approaches, in particular, with contextuality. Indeed, I believe that
contextuality was rather foreign to Itamar Pitowsky’s thinking: the term “contex-
tuality” appears marginally – as in “a different context” – in his book Quantum
Probability – Quantum Logic (Pitowsky 1989b), nowhere in his reviews on Boole-
Bell type inequalities (Pitowsky 1989a, 1994), and mostly with reference to
contextual quantum probabilities in his late writings (Pitowsky 2006). The emphasis
on value indefiniteness/indeterminacy was, I believe, independently shared by Asher
Peres as well as Ernst Specker.
I met Itamar Pitowsky (Bub and Demopoulos 2010) personally rather late; after
he gave a lecture entitled “All Bell Inequalities” in Vienna (ESI – The Erwin
Schrödinger International Institute for Mathematical Physics 2001) on September
6th, 2000. Subsequent discussions resulted in a joint paper (Pitowsky and Svozil
2001) (stimulating further research (Sliwa 2003; Colins and Gisin 2004)). It presents
an application of his correlation polytope method (Pitowsky 1986, 1989a,b, 1991,
1994) to more general configurations than had been studied before. Thereby semi-
automated symbolic as well as numeric computations have been used.
Nevertheless, the violations of what Boole called (Boole 1862, p. 229) “condi-
tions of possible experience,” obtained through solving the hull problem of classical
correlation polytopes, was just one route to quantum indeterminacy pursued by
Itamar Pitowsky. One could identify at least two more passages he contributed
to: One approach (Pitowsky 2003, 2006) compares differences of classical with
quantum predictions through conditions and constraints imposed by certain inter-
twined configurations of observables which I like to call quantum clouds (Svozil
2017b). And another approach (Pitowsky 1998; Hrushovski and Pitowsky 2004)
pushes these predictions to the limit of logical inconsistency; such that any attempt
of a classical description fails relative to the assumptions. In what follows we shall
follow all three pursuits and relate them to new findings.
24 Roots and (Re)sources of Value (In)definiteness Versus Contextuality 523

24.2 Stochastic Value Indefiniteness/Indeterminacy by

Boole-Bell Type Conditions of Possible Experience

The basic idea to obtain all classical predictions – including classical probabilities,
expectations as well as consistency constraints thereof – associated with (mostly
complementary; that is, non-simultaneously measurable) collections of observables
is quite straightforward: Figure out all “extreme” cases or states which would be
classically allowed. Then construct all classically conceivable situations by forming
suitable combinations of the former.
Formally this amounts to performing the following steps (Pitowsky 1986,
1989a,b, 1991, 1994):
• Contemplate about some concrete structure of observables and their interconnec-
tions in intertwining observables – the quantum cloud.
• Find all two-valued states of that quantum cloud. (In the case of “contextual
inequalities” (Cabello 2008) include all variations of true/1 and false/0, irre-
spective of exclusivity; thereby often violating the Kolmogorovian axioms of
probability theory even within a single context.)
• Depending on one’s preferences, form all (joint) probabilities and expectations.
• For each of these two-valued states, evaluate the joint probabilities and expec-
tations as products of the single particle probabilities and expectations they are
formed of (this reflects statistical independence of the constituent observables).
• For each of the two-valued states, form a tuple containing these relevant (joint)
probabilities and expectations.
• Interpret this tuple as a vector.
• Consider the set of all such vectors – there are as many as there are two-
valued states, and their dimension depends on the number of (joint) probabilities
and expectations considered – and interpret them as vertices forming a convex
polytope.
• The convex combination of all conceivable two-valued states yields the surface
of this polytope; such that every point inside its convex hull corresponds to a
classical probability distribution.
• Determine the conditions of possible experience by solving the hull problem
– that is, by computing the hyperplanes which determine the inside–versus–
outside criteria for that polytope. These then can serve as necessary criteria for
all classical probabilities and expectations considered.
The systematic application of this method yields necessary criteria for classi-
cal probabilities and expectations which are violated by the quantum probabil-
ities and expectations. Since I have reviewed this subject exhaustively (Svozil
2018c, Sect. 12.9) (see also Svozil 2017a) I have just sketched it to obtain
a taste for its relevance for quantum indeterminacy. As is often the case in
mathematical physics the method seems to have been envisioned independently
a couple of times. From its (to the best of my knowledge) inception by Boole
(1862) it has been discussed in the measure theoretic context by Chochet the-
524 K. Svozil

ory (Bishop and Leeuw 1959) and by Vorob’ev (1962). Froissart (Froissart 1981;
Cirel’son (=Tsirel’son) 1993) might have been the first explicitly proposing it
as a method to generalized Bell-type inequalities. I suggested its usefulness for
non-Boolean cases (Svozil 2001) with “enough” two-valued states; preferable
sufficiently many to allow a proper distinction/separation of all observables (cf.
Kochen and Specker’s Theorem 0 (Kochen and Specker 1967, p. 67)). Consider-
ation of the pentagon/pentagram logic – that is, five cyclically intertwined con-
texts/blocks/Boolean subalgebras/cliques/orthonormal bases popularized the sub-
ject and also rendered new predictions which could be used to differentiate classical
from quantized systems (Klyachko 2002; Klyachko et al. 2008; Bub and Stairs 2009,
2010; Badzia̧g et al. 2011).
A caveat: the obtained criteria involve multiple mutually complementary sum-
mands which are not all simultaneously measurable. Therefore, different terms,
when evaluated experimentaly, correspond to different, complementary measure-
ment configurations. They are obtained at different times and on different particles
and samples.
Explicit, worked examples can, for instance, be found in Pitowsky’s
book (Pitowsky 1989b, Section 2.1), or papers (Pitowsky 1994) (see also Froissart’s
example (Froissart 1981)). Empirical findings are too numerous to even attempt
a just appreciation of all the efforts that went into testing classicality. There is
overwhelming evidence that the quantum predictions are correct; and that they
violate Boole’s conditions of possible classical experience (Clauser 2002) relative
to the assumptions (basically non-contextual realism and locality).
So, if Boole’s conditions of possible experience are violated, then they can no
longer be considered appropriate for any reasonable ontology forcing “reality”
upon them. This includes the realistic (Stace 1934) existence of hypothetical
counterfactual observables: “unperformed experiments seem to have no consis-
tent outcomes” (Peres 1978). The inconsistency of counterfactuals (in Specker’s
scholastic terminology infuturabilities (Specker 1960, 2009)) provides a connection
to value indefiniteness/indeterminacy – at least, and let me again repeat earlier
provisos, relative to the assumptions. More of this, piled higher and deeper, has
been supplied by Itamar Pitowsky, as will be discussed later.

24.3 Interlude: Quantum Probabilities from Pythagorean

“Views on Vectors”

Quantum probabilities are vector based. At the same time those probabilities mimic
“classical” ones whenever they must be classical; that is, among mutually commut-
ing observables which can be measured simultaneously/concurrently on the same
particle(s) or samples – in particular, whenever those observables correspond to
projection operators which are either orthogonal (exclusive) or identical (inclusive).
24 Roots and (Re)sources of Value (In)definiteness Versus Contextuality 525

At the same time, quantum probabilities appear “contextual” (I assume he had

succumbed to the prevalent nomenclature at that late time) according to Itamar
Pitowsky’s late writings (Pitowsky 2006) if they need not be classical: namely
among non-commuting observables. (The term “needs not” derives its justification
from the finding that there exist situations (Moore 1956; Wright 1990) involv-
ing complementary observables with a classical probability interpretation (Svozil
2005)).
Thereby, classical probability theory is maintained for simultaneously co-
measurable (that is, non-complementary) observables. This essentially amounts
to the validity of the Kolmogorov axioms of probability theory of such observables
within a given context/block/Boolean subalgebra/clique/orthonormal basis,
whereby the probability of an event associated with an observable
• is a non-negative real number between 0 and 1;
• is 1 for an event associated with an observable occurring with certainty (in
particular, by considering any observable or its complement); as well as
• additivity of probabilities for events associated with mutually exclusive observ-
ables.
Sufficiency is assured by an elementary geometric argument (Gleason 1957)
which is based upon the Pythagorean theorem; and which can be used to explicitly
construct vector-based probabilities satisfying the aforementioned Kolmogorov
axioms within contexts: Suppose a pure state of a quantized system is formalized by
the unit state vector |ψ. Consider some orthonormal basis B = {|e √ 1 , . . . , |en }
of V. Then the square Pψ (ei ) = |ψ|ei |2 of the length/norm ψ|ei ei |ψ
of the orthogonal projection (ψ|ei ) |ei  of that unit vector |ψ along the basis
element |ei  can be interpreted as the probability of the event associated with the
0 − 1-observable (proposition) associated with the basis vector |ei  (or rather the
orthogonal projector Ei = |ei ei | associated with the dyadic product of the basis
vector |ei ); given a quantized physical system which  has been prepared to be in
a pure state |ψ. Evidently, 1 ≤ Pψ (ei ) ≤ 1, and ni=1 Pψ (ei ) = 1. In that
Pythagorean way, every context, formalized by an orthonormal basis B, “grants
a (probabilistic) view” on the pure state |ψ.
It can be expected that these Pythagorean-style probabilities are different from
classical probabilities almost everywhere – that is, for almost all relative measure-
ment positions. Indeed, for instance, whereas classical two-partite correlations are
linear in the relative measurement angles, their respective quantum correlations
follow trigonometric functions – in particular, the cosine for “singlets” (Peres 1993).
These differences, or rather the vector-based Pythagorean-style quantum probabil-
ities, are the “root cause” for violations of Boole’s aforementioned conditions of
possible experience in quantum setups.
Because of the convex combinations from which they are derived, all of these
conditions of possible experience contain only linear constraints (Boole 1854,
1862; Fréchet 1935; Hailperin 1965, 1986; Ursic 1984, 1986, 1988; Beltrametti
and Maçzyński 1991, 1993, 1994, 1995; Pykacz and Santos 1991; Sylvia and
Majernik 1992; Dvurečenskij and Länger 1994; Beltrametti et al. 1995; Del Noce
526 K. Svozil

1995; Länger and Maçzyński 1995; Dvurečenskij and Länger 1995a,b; Beltrametti
and Bugajski 1996; Pulmannová 2002). And because linear combinations of linear
operators remain linear, one can identify the terms occurring in conditions of possi-
ble experience with linear self-adjoint operators, whose sum yields a self-adjoint
operator, which stands for the “quantum version” of the respective conditions
of possible experience. This operator has a spectral decomposition whose min-
max eigenvalues correspond to the quantum bounds (Filipp and Svozil 2004a,b),
which thereby generalize the Tsirelson bound (Cirel’son (=Tsirel’son) 1980). In
that way, every condition of possible experience which is violated by the quantum
probabilities provides a direct criterium for non-classicality.

24.4 Classical Value Indefiniteness/Indeterminacy by Direct

Observation

In addition to the “fragmented, explosion view” criteria allowing “nonlocality” via

Einstein separability (Weihs et al. 1998) among its parts, classical predictions from
quantum clouds – essentially intertwined (therefore the Hilbert space dimensionality
has to be greater than two) arrangements of contexts – can be used as a criterium for
quantum advantage over (or rather “otherness” or “distinctiveness” from) classical
predictions. Thereby it is sufficient to observe of a single outcome of a quantized
system which directly contradicts the classical predictions.
One example of such a configuration of quantum observables forcing a “one-
zero rule” (Svozil 2009b) because of a true-implies-false set of two-valued classical
states (TIFS) (Cabello et al. 2018) is the “Specker bug” logic (Kochen and Specker
1965, Fig. 1, p. 182) called “cat’s cradle” (Pitowsky 2003, 2006) by Itamar Pitowsky
(see also Belinfante (1973, Fig. B.l. p. 64), Stairs (1983, p. 588–589), Clifton (1993,
Sects. IV, Fig. 2) and Pták and Pulmannová (1991, p. 39, Fig. 2.4.6) for early
discussions), as depicted in Fig. 24.1.
For such configurations, it is often convenient to represent both its labels as well
as the classical probability distributions in terms of a partition logic (Svozil 2005)
of the set of two-valued states – in this case, there are 14 such classical states. Every
maximal observable is characterized by a context. The atoms of this context are
labeled according to the indices of the two-valued measure with the value 1 on this
atom. The axioms of probability theory require that, for each two-valued state, and
within each context, there is exactly one such atom. As a result, as long as the set of
two-valued states is separating (Kochen and Specker 1967, Theorem 0), one obtains
a set of partitions of the set of two-valued states; each partition corresponding to a
context.
Classically, if one prepares the system to be in the state {1, 2, 3} – standing for
any one of the classical two-valued states 1, 2 or 3 or their convex combinations –
then there is no chance that the “remote” target state {7, 10, 13} can be observed.
A direct observation of quantum advantages (or rather superiority in terms of the
24 Roots and (Re)sources of Value (In)definiteness Versus Contextuality 527

{3, 9, 13, 14}

{5, 7, 8, 10, 11} {1, 2, 4, 6, 12}

{4, 6, 9, 12, 13, 14} {3, 5, 8, 9, 11, 14}

{1, 4,
{1, 2, 3} 5, 10, {7, 10, 13}
11, 12}

{4, 5, 6, 7, 8, 9} {2, 6, 8, 11, 12, 14}

{10, 11, 12, 13, 14} {1, 3, 4, 5, 9}

{2, 6, 7, 8}

Fig. 24.1 The convex structure of classical probabilities in this (Greechie) orthogonality diagram
representation of the Specker bug quantum or partition logic is reflected in its partition logic,
obtained through indexing all 14 two-valued measures, and adding an index 1 ≤ i ≤ 14 if the
ith two-valued measure is 1 on the respective atom. Concentrate on the outermost left and right
observables, depicted by squares: Positivity and convexity requires that 0 ≤ λi ≤ 1 and λ1 + λ2 +

λ3 + λ7 + λ10 + λ13 ≤ 14 i=1 λi = 1. Therefore, if a classical system is prepared (a generalized
urn model/automaton logic is “loaded”) such that λ1 + λ2 + λ3 = 1, then λ7 + λ10 + λ13 = 0,
which results in a TIFS: the classical prediction is that the latter outcome never occurs if the former
preparation is certain

frequencies predicted with respect to classical frequencies) is then suggested by

some faithful orthogonal representation (FOR) (Lovász et al. 1989; Parsons and
Pisanski 1989; Cabello et al. 2010; Solís-Encina and Portillo 2015) of this graph.
In the particular Specker bug/cats cradle configuration, an elementary geometric
argument (Cabello 1994, 1996) forces the relative angle between the quantum  states
√ 
|{1, 2, 3} and |{7, 10, 13} in three dimensions to be not smaller than arctan 2 2 ,
so that the quantum prediction of the occurrence of the event associated with state
|{7, 10, 13}, if the system was prepared in state |{1,
 2, 3} √is that
 the probability
can be at most |{1, 2, 3}|{7, 10, 13}| = cos arctan 2 2
2 2 = 19 . That is,
on the average, if the system was prepared in state |{1, 2, 3} at most one of 9
outcomes indicates that the system has the property associated with the observable
|{7, 10, 13}|{7, 10, 13}|. The occurrence of a single such event indicates quantum
advantages over the classical prediction of non-occurrence.
This limitation is only true for the particular quantum cloud involved. Similar
arguments with different quantum clouds resulting in TIFS can be extended to
arbitrary small relative angles between preparation and measurement states, so
that the relative quantum advantage can be made arbitrarily high (Abbott et al.
2015; Ramanathan et al. 2018). Classical value indefiniteness/indeterminacy comes
naturally: because – at least relative to the assumptions regarding non-contextual
value definiteness of truth assignments, in particular, of intertwining, observables
– the existence of such definite values would enforce non-occurrence of outcomes
which are nevertheless observed in quantized systems.
528 K. Svozil

Very similar arguments against classical value definiteness can be inferred from
quantum clouds with true-implies-true sets of two-valued states (TITS) (Stairs 1983;
Clifton 1993; Johansen 1994; Vermaas 1994; Belinfante 1973; Pitowsky 1982;
Hardy 1992, 1993; Boschi et al. 1997; Cabello and García-Alcaine 1995; Cabello
et al. 1996, 2013, 2018; Cabello 1997; Badzia̧g et al. 2011; Chen et al. 2013).
There the quantum advantage is in the non-occurrence of outcomes which classical
predictions mandate to occur.

24.5 Classical Value Indefiniteness/Indeterminacy Piled

Higher and Deeper: The Logical Indeterminacy
Principle

For the next and final stage of classical value indefiniteness/indeterminacy on

quantum clouds (relative to the assumptions) one can combine two logics with
simultaneous classical TIFS and TITS properties at the same terminals. That is,
suppose one is preparing the same “initial” state, and measuring the same “target”
observable; nevertheless, contemplating the simultaneous counterfactual existence
of two different quantum clouds of intertwined contexts interconnecting those
fixated “initial” state and measured “target” observable. Whenever one cloud has
the TIFS and another cloud the TITS property (at the same terminals), those
quantum clouds induce contradicting classical predictions. In such a setup the
only consistent choice (relative to the assumptions; in particular, omni-existence
and context independence) is to abandon classical value definiteness/determinacy.
Because the assumption of classical value definiteness/determinacy for any such
logic, therefore, yields a complete contradiction, thereby eliminating prospects for
hidden variable models (Abbott et al. 2012, 2015; Svozil 2017b) satisfying the
assumptions.
Indeed, suppose that a quantized system is prepared in some pure quantum
state. Then Itamar Pitowsky’s (Pitowsky 1998; Hrushovski and Pitowsky 2004)
indeterminacy principle states that – relative to the assumptions; in particular,
global classical value definiteness for all observables involved, as well as context-
independence of observables in which contexts intertwine – any other distinct
(non-collinear) observable which is not orthogonal can neither occur nor not occur.
This can be seen as an extension of both Gleason’s theorem (Gleason 1957; Zierler
and Schlessinger 1965) as well as the Kochen-Specker theorem (Kochen and
Specker 1967) implying and utilizing the non-existence of any two-valued global
truth assignments on even finite quantum clouds.
For the sake of a concrete example consider the two TIFS and TITS clouds –
that is, logics with 35 intertwined binary observables (propositions) in 24 contexts
– depicted in Fig. 24.2 (Svozil 2018b). They represent quantum clouds with the
same terminal points {1} ≡ {1 } and {2, 3, 4, 5, 6, 7} ≡ {1 , 2 , 3 , 4 , 5 }, forcing
24 Roots and (Re)sources of Value (In)definiteness Versus Contextuality 529

{1}
a

{2,3,4,5,6,7}

b
{1′}

{1′,2′,3′,4′,5′}

c 4 36 º {} 5

12 31
15 13
30 14 29
10 17
28 20 11
7 33 23
34 16
27 21
6 25
1 26
24 35
18 32
22 9

19 8

37 º {1″,2″,3″,4″}
2 3

Fig. 24.2 (a) TIFS cloud, and (b) TITS cloud with only a single overlaid classical value
assignment if the system is prepared in state |1 (Svozil 2018b). (c) The combined cloud from
(a) and (b) has no value assignment allowing 36 = {} to be true/1; but still allows 8 classical value
assignments enumerated by Table 24.1, with overlaid partial coverage common to all of them. A
faithful orthogonal realization is enumerated in Abbott et al. (2015, Table. 1, p. 102201–7)
530 K. Svozil

the latter ones (that is, {2, 3, 4, 5, 6, 7} and {1 , 2 , 3 , 4 , 5 }) to be false/0 and true/1,
respectively, if the former ones (that is, {1} ≡ {1 }) are true/1.
Formally, the only two-valued states on the logics depicted in Fig. 24.2a and
b which allow v({1}) = v  ({1 }) = 1 requires that v({2, 3, 4, 5, 6, 7}) = 0
but v  ({1 , 2 , 3 , 4 , 5 }) = 1 − v({2, 3, 4, 5, 6, 7}), respectively. However, both
these logics have a faithful orthogonal representation (Abbott et al. 2015, Table. 1,
p. 102201–7) in terms of vectors which coincide in |{1} = |{1 }, as well as
in |{2, 3, 4, 5, 6, 7} = |{1 , 2 , 3 , 4 , 5 }, and even in all of the other adjacent
observables.
The combined logic, which features 37 binary observables (propositions) in 26
contexts has no longer a classical interpretation in terms of a partition logic, as the
8 two-valued states enumerated in Table 24.1 cannot mutually separate (Kochen
and Specker 1967, Theorem 0) the observables 2, 13, 15, 16, 17, 25, 27 and 36,
respectively.
It might be amusing to keep in mind that, because of non-separability (Kochen
and Specker 1967, Theorem 0) of some of the binary observables (propositions),
there does not exist a proper partition logic. However, there exist generalized
urn (Wright 1978, 1990) and finite automata (Moore 1956; Schaller and Svozil
1995, 1996) model realisations thereof: just consider urns “loaded” with balls which
have no colored symbols on them; or no such balls at all, for the binary observables
(propositions) 2, 13, 15, 16, 17, 25, 27 and 36. In such cases it is no more possible to
empirically reconstruct the underlying logic; yet if an underlying logic is assumed
then – at least as long as there still are truth assignments/two-valued states on
the logic – “reduced” probability distributions can be defined, urns can be loaded,
and automata prepared, which conform to the classical predictions from a convex
combination of these truth assignments/two-valued states – thereby giving rise to
“reduced” conditions of experience via hull computations.
For global/total truth assignments (Pitowsky 1998; Hrushovski and Pitowsky
2004) as well as for local admissibility rules allowing partial (as opposed to
total, global) truth assignments (Abbott et al. 2012, 2015), such arguments can be
extended to cover all terminal states which are neither collinear nor orthogonal.
One could point out that, insofar as a fixed state has to be prepared the resulting
value indefiniteness/indeterminacy is state dependent. One may indeed hold that
the strongest indication for quantum value indefiniteness/indeterminacy is the total
absence/non-existence of two-valued states, as exposed in the Kochen-Specker
theorem (Kochen and Specker 1967). But this is rather a question of nominalistic
taste, as both cases have no direct empirical testability; and as has already been
pointed out by Clifton in a private conversation in 1995: “how can you measure a
contradiction?”
Table 24.1 Enumeration of the 8 two-valued states on 37 binary observables (propositions) of the combined quantum clouds/logics depicted in Fig. 24.2a
and b. Row vector indicate the state values on the observables, column vectors the values on all states per the respective observable
# 1 2 3 4 ··· ··· ··· 34 35 36 37
1 1 0 0 1 0 0 0 0 0 0 1 1 0 0 0 0 0 0 1 0 1 0 0 1 0 1 0 0 1 1 0 1 1 1 1 0 1
2 1 0 0 1 0 0 0 0 0 0 1 1 0 0 0 0 0 0 1 1 0 0 1 1 0 0 0 0 0 1 1 1 1 1 1 0 1
3 1 0 0 0 1 0 0 0 0 1 0 0 0 1 0 0 0 0 1 0 1 0 0 1 0 1 0 1 0 0 1 1 1 1 1 0 1
4 1 0 0 0 1 0 0 0 0 1 0 0 0 1 0 0 0 1 0 0 1 1 0 0 0 1 0 0 0 1 1 1 1 1 1 0 1
5 1 0 1 1 0 1 0 1 0 0 1 1 0 0 0 0 0 1 0 1 0 0 0 1 0 1 0 0 1 1 0 0 0 1 1 0 0
6 1 0 1 1 0 1 0 0 1 0 0 1 0 1 0 0 0 1 0 1 0 0 0 1 0 1 0 0 1 1 0 1 0 1 0 0 0
7 1 0 1 0 1 1 0 1 0 1 0 0 0 1 0 0 0 1 0 1 0 0 0 1 0 1 0 1 0 0 1 0 0 1 1 0 0
8 1 0 1 0 1 0 1 1 0 0 0 1 0 1 0 0 0 1 0 1 0 0 0 1 0 1 0 1 0 0 1 0 1 0 1 0 0
24 Roots and (Re)sources of Value (In)definiteness Versus Contextuality
531
532 K. Svozil

24.6 The “Message” of Quantum (In)determinacy

At the peril of becoming, as expressed by Clauser (2002), “evangelical,” let me

“sort things out” from my own very subjective and private perspective. (Readers
adverse to “interpretation” and the semantic, “meaning” aspects of physical theory
may consider stop reading at this point.)
Thereby one might be inclined to follow Planck (against Feynman (Clauser
2002; Mermin 1989a,b)) and hold it as being not too unreasonable to take
scientific comprehensibility, rationality, and causality as a (Planck 1932, p. 539)
(see also Earman 2007, p. 1372) “heuristic principle, a sign-post . . . to guide us
in the motley confusion of events and to show us the direction in which scientific
research must advance in order to attain fruitful results.”
So what does all of this – the Born rule of quantum probabilities and its deriva-
tion by Gleason’s theorem from the Kolmogorovian axioms applied to mutually
comeasurable observables, as well as its consequences, such as the Kochen-Specker
theorem, the plethora of violations of Boole’s conditions of possible experience,
Pitowsky’s indeterminacy principle and more recent extensions and variations
thereof – “try to tell us?”
First, observe that all of the aforementioned postulates and findings are (based
upon) assumptions; and thus consequences of the latter. Stated differently, these
findings are true not in the absolute, ontologic but in the epistemic sense: they hold
relative to the axioms or assumptions made.
Thus, in maintaining rationality one needs to grant oneself – or rather one is
forced to accept – the abandonment of at least some or all assumptions made. Some
options are exotic; for instance, Itamar Pitowsky’s suggestions to apply paradoxical
set decompositions to probability measures (Pitowsky 1983, 1986). Another “exotic
escape option” is to allow only unconnected (non-intertwined) contexts whose
observables are dense (Godsil and Zaks 1988, 2012; Meyer 1999; Havlicek et al.
2001). Some possibilities to cope with the findings are quite straightforward, and
we shall concentrate our further attention to those (Svozil 2009b).

24.6.1 Simultaneous Definiteness of Counterfactual,

Complementary Observables, and Abandonment of
Context Independence

Suppose one insists on the simultaneous definite omni-existence of mutually

complementary, and therefore necessarily counterfactual, observables. One straight-
forward way to cope with the aforementioned findings is the abandonment of
context-independence of intertwining observables.
There is no indication in the quantum formalism which would support such an
assumption, as the respective projection operators do not in any way depend on the
contexts involved. However, one may hold that the outcomes are context dependent
24 Roots and (Re)sources of Value (In)definiteness Versus Contextuality 533

as functions of the initial state and the context measured (Svozil 2009a, 2012;
Dzhafarov et al. 2017); and that they actually “are real” and not just “idealistically
occur in our imagination;” that is, being “mental through-and-through” (Segal
and Goldschmidt 2017/2018). Early conceptualizations of context-dependence aka
contextuality can be found in Bohr’s remark (in his typical Nostradamus-like
style) (Bohr 1949) on “the impossibility of any sharp separation between the
behavior of atomic objects and the interaction with the measuring instruments which
serve to define the conditions under which the phenomena appear.” Bell, referring
to Bohr, suggested (Bell 1966), Sec. 5) that “the result of an observation may
reasonably depend not only on the state of the system (including hidden variables)
but also on the complete disposition of the apparatus.”
However, the common, prevalent, use of the term “contextuality” is not an
explicit context-dependent form, as suggested by the realist Bell in his earlier quote,
but rather a situation where the classical predictions of quantum clouds are violated.
More concretely, if experiments on quantized systems violate certain Boole-Bell
type classical bounds or direct classical predictions, the narratives claim to have
thereby “proven contextuality” (e.g., see Hasegawa et al. (2006), Cabello et al.
(2008), Cabello (2008), Bartosik et al. (2009), Amselem et al. (2009), Bub and Stairs
(2010) and Cabello et al. (2013) for a “direct proof of quantum contextuality”).
What if we take Bell’s proposal of a context dependence of valuations –
and consequently, “classical” contextual probability theory – seriously? One of
the consequences would be the introduction of an uncountable multiplicity of
counterfactual observables. An example to illustrate this multiplicity – comparable
to de Witt’s view of Everett’s relative state interpretation (Everett III 1973) – is
the uncountable set of orthonormal bases of R3 which are all interconnected at the
same single intertwining element. A continuous angular parameter characterizes the
angles between the other elements of the bases, located in the plane orthogonal to
that common intertwining element. Contextuality suggests that the value assignment
of an observable (proposition) corresponding to this common intertwining element
needs to be both true/1 and false/0, depending on the context involved, or whenever
some quantum cloud (collection of intertwining observables) demands this through
consistency requirements.
Indeed, the introduction of multiple quantum clouds would force any context
dependence to also implicitly depend on this general perspective – that is, on the
respective quantum cloud and its faithful orthogonal realization, which in turn
determines the quantum probabilities via the Born-Gleason rule: Because there exist
various different quantum clouds as “pathways interconnecting” two observables,
context dependence needs to vary according to any concrete connection between
the prepared and the measured state.
A single context participates in an arbitrary, potentially infinite, multiplicity of
quantum clouds. This requires this one context to “behave very differently” when
it comes to contextual value assignments. Alas, as quantum clouds are hypothetical
constructions of our mind and therefore “mental through-and-through” (Segal and
Goldschmidt 2017/2018), so appears context dependence: as an idealistic concept,
534 K. Svozil

devoid of any empirical evidence, created to rescue the desideratum of omni-

realistic existence.
Pointedly stated, contextual value assignments appear both utterly ad hoc and
abritrary – like a deus ex machina “saving” the desideratum of a classical omni-
value definite reality, whereby it must obey quantum probability theory without
grounding it (indeed, in the absence of any additional criterium or principle there
is no reason to assume that the likelihood of true/1 and false/0 is other than
50:50); as well as highly discontinuous. In this latter, discontinuity respect, context
dependence is similar to the earlier mentioned breakup of the intertwine observables
by reducing quantum observables to disconnected contexts (Godsil and Zaks 1988,
2012; Meyer 1999; Havlicek et al. 2001).
It is thereby granted that these considerations apply only to cases in which
the assumptions of context independence are valid throughout the entire quantum
cloud – that is, uniformly: for all observables in which contexts intertwine. If this
were not the case – say, if only a single one observable occurring in intertwining
contexts is allowed to be context-dependent (Svozil 2012; Simmons 2020) – the
respective clouds taylored to prove Pitowsky’s Logical Indeterminacy Principle and
similar, as well as the Kochen-Specker theorems do not apply; and therefore the
aforementioned consequences are invalid.

24.6.2 Abandonment of Omni-Value Definiteness of

Observables in All But One Context

Nietzsche once speculated (Nietzsche 1887, 2009–,,) that what he has called
“slave morality” originated from superficially pretending that – in what later Blair
(aka Orwell 1949) “doublespeak” – weakness means strength. In a rather similar
sense the lack of comprehension – Planck’s “sign-post” – and even the resulting
inconsistencies tended to become reinterpreted as an asset: nowadays consequences
of the vector-based quantum probability law are marketed as “quantum supremacy”
– a “quantum magic” or “hocus-pocus” (Svozil 2016) of sorts.
Indeed, future centuries may look back at our period, and may even call it a
second “renaissance” period of scholasticism (Specker 1960). In years from now
historians of science will be amused about our ongoing queer efforts, the calamities
and “magic” experienced through our painful incapacity to recognize the obvious –
that is, the non-existence and therefore value indefiniteness/indeterminacy of certain
counterfactual observables – namely exactly those mentioned in Itamar Pitowsky’s
indeterminacy principle.
This principle has a positive interpretation of a quantum state, defined as the
maximal knowledge obtainable by simultaneous measurements of a quantized
system; or, conversely, as the maximal information content encodable therein. This
can be formalized in terms of the value definiteness of a single (Zeilinger 1999;
Svozil 2002, 2004, 2018b; Grangier 2002) context – or, in a more broader (non-
24 Roots and (Re)sources of Value (In)definiteness Versus Contextuality 535

operational) perspective, the continuum of contexts intertwined by some prepared

pure quantum state (formalized as vector or the corresponding one-dimensional
orthogonal projection operator). In terms of Hilbert space quantum mechanics this
amounts to the claim that the only value definite entity can be a single orthonormal
basis/maximal operator; or a continuum of maximal operators whose spectral
sum contain proper “true intertwines.” All other “observables” grant an, albeit
necessarily stochastic, value indefinite/indeterministic, view on this state.
If more than one context is involved we might postulate that all admissable
probabilities should at least satisfy the following criterium: they should be classical
Kolmogorov-style within any single particular context (Gleason 1957). It has been
suggested (Aufféves and Grangier 2017, 2018) that this can be extended and
formalized in a quantum multi-context environment by a double stochastic matrix
whose entries P (ei , fj ), with 1 ≤ i, j ≤ n (n is the number of distinct “atoms” or
exclusive outcomes in each context) are identified by the conditional probabilities of
one atom fj in the second context, relative to a given one atom ei in the first context.
The general multi-context case yields row stochastic matrices (Svozil 2018a).
Various types of decompositions of those matrices exist for particular cases:
• By the Birkhoff-von Neumann theorem double stochastic matrices can be
represented by the Birkhoff polytope spanned by the convex k hull of the set
of permutation matrices: let λ1 , . . . , λ k ≥ 0 such that l=1 λl = 1, then
k
P (ei , fj ) = l=1 λl l . Since there exist n! permutations of n elements, k
ij
will be bounded from above by k ≤ n!. Note that this type of decomposition 1 may2
not be unique, as the space spanned the permutation matrices is (n − 1)2 + 1 -
dimensional; with n! > (n − 1)2 + 1 for n > 2. Therefore, the bound from above
can be improved such that decompositions with k ≤ (n − 1)2 + 1 = n2 − 2(n + 1)
exist (Marcus and Ree 1959). Formally, a permutation matrix has a quasi-
vectorial (Mermin 2007)  decomposition in terms of the canonical (Cartesian)
n
basis, such that, i = j =1 |ej eπ i (j ) |, where |ej  represents the n-tuple
associated with the j th basis vector of the canonical (Cartesian) basis, and π i (j )
stands for the ith permutation of j .
• Vector based probabilities allow the following  decomposition
 (Aufféves and
Grangier 2017, 2018): P (ei , fj ) = Trace Ei RFj R , where Ei and Fi are
elements of contexts, formalized by two sets of mutually orthogonal projection
operators, and R is a real positive diagonal 2
  matrix such that the trace of R
equals the dimension n, and Trace Ei R 2 = 1. The quantum mechanical
Born rule is recovered
 by
 identifying R = I n with the identity matrix, so that
P (ei , fj ) = Trace Ei Fj .
• There exist more “exotic” probability measures on “reduced” proposi-
tional spaces such as Wright’s 2-state dispersion-free measure on the
pentagon/pentagram (Wright 1978), or another type of probability measure
based on a discontinuous 3(2)-coloring of the set of all unit vectors with rational
coefficients (Godsil and Zaks 1988, 2012; Meyer 1999; Havlicek et al. 2001)
whose decomposition appear to be ad hoc; at least for the time being.
536 K. Svozil

Where might this aforementioned type of stochasticism arise from? It could well
be that it is introduced by interactions with the environment; and through the many
uncontrollable and, for all practical purposes (Bell 1990), huge number of degrees
of freedom in unknown states.
The finiteness of physical resources needs not prevent the specification of a
particular vector or context. Because any other context needs to be operationalized
within the physically feasible means available to the respective experiment: it is the
measurable coordinate differences which count; not the absolute locatedness relative
to a hypothetical, idealistic absolute frame of reference which cannot be accessed
operationally.
Finally, as the type of context envisioned to be value definite can be expressed in
terms of vector spaces equipped with a scalar product – in particular, by identifying
a context with the corresponding orthonormal basis or (the spectral decomposition
of) the associated maximal observable(s) – one may ask how one could imagine
the origin of such entities? Abstractly vectors and vector spaces could originate
from a great variety of very different forms; such as from systems of solutions
of ordinary linear differential equations. Any investigation into the origins of the
quantum mechanical Hilbert space formalism itself might, if this turns out to be a
progressive research program (Lakatos 1978), eventually yield to a theory indicating
operational physical capacities beyond quantum mechanics.

24.7 Biographical Notes on Itamar Pitowsky

I am certainly not in the position to present a view of Itamar Pitowsky’s thinking.

Therefore I shall make a few rather anecdotal observations. First of all, he seemed
to me as one of the most original physicists I have ever met – but that might
be a triviality, given his opus. One thing I realized was that he exhibited a –
sometimes maybe even unconscious, but sometimes very outspoken – regret that
he was working in a philosophy department. I believe he considered himself
rather a mathematician or theoretical physicist. To this I responded that being in a
philosophy department might be rather fortunate because there one could “go wild”
in every direction; allowing much greater freedom than in other academic realms.
But, of course, this had no effect on his uneasiness.
He was astonished that I spent a not so little money (means relative to my
investment capacities) in an Israeli internet startup company which later flopped,
depriving me of all but a fraction of what I had invested. He told me that, at least
at that point, many startups in Israel had been put up intentionally only to attract
money from people like me; only to collapse later.
A late project of his concerned quantum bounds in general; maybe in a similar
– graph theoretical and at the time undirected to quantum – way as Grötschel,
Lovász and Schrijver’s theta body (Grötschel et al. 1986; Cabello et al. 2014). The
idea was not just deriving absolute (Cirel’son (=Tsirel’son) 1980) or parameterized,
continuous (Filipp and Svozil 2004a,b) bounds for existing classical conditions of
24 Roots and (Re)sources of Value (In)definiteness Versus Contextuality 537

possible experience obtained by hull computations of polytopes; but rather genuine

quantum bounds on, say, Einstein-Podolsky-Rosen type setups.

Acknowledgements I kindly acknowledge enlightening criticism and suggestions by Andrew

W. Simmons, as well as discussions with Philippe Grangier on the characterization of quantum
probabilities. All remaining misconceptions and errors are mine. The author acknowledges the
support by the Austrian Science Fund (FWF): project I 4579-N and the Czech Science Foundation:
project 20-09869L.

References

Abbott, A. A., Calude, C. S., Conder, J., & Svozil, K. (2012). Strong Kochen-Specker theorem
and incomputability of quantum randomness. Physical Review A, 86, 062109. https://doi.org/
10.1103/PhysRevA.86.062109, arXiv:1207.2029
Abbott, A. A., Calude, C. S., & Svozil, K. (2014) Value-indefinite observables are almost
everywhere. Physical Review A, 89, 032109. https://doi.org/10.1103/PhysRevA.89.032109,
arXiv:1309.7188
Abbott, A. A., Calude, C. S., & Svozil, K. (2015). A variant of the Kochen-Specker theorem
localising value indefiniteness. Journal of Mathematical Physics, 56(10), 102201. https://doi.
org/10.1063/1.4931658, arXiv:1503.01985
Amselem, E., Rådmark, M., Bourennane, M., & Cabello, A. (2009). State-independent quantum
contextuality with single photons. Physical Review Letters, 103(16), 160405. https://doi.org/
10.1103/PhysRevLett.103.160405
Aufféves, A., & Grangier, P. (2017). Recovering the quantum formalism from physically
realist axioms. Scientific Reports, 7(2123), 43365 (1–9). https://doi.org/10.1038/srep43365,
arXiv:1610.06164
Aufféves, A., & Grangier, P. (2018). Extracontextuality and extravalence in quantum mechanics.
Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering
Sciences, 376(2123), 20170311. https://doi.org/10.1098/rsta.2017.0311, arXiv:1801.01398
Badzia̧g, P., Bengtsson, I., Cabello, A., Granström, H., & Larsson, J. A. (2011). Pentagrams and
paradoxes. Foundations of Physics, 41. https://doi.org/10.1007/s10701-010-9433-3.
Bartosik, H., Klepp, J., Schmitzer, C., Sponar, S., Cabello, A., Rauch, H., & Hasegawa, Y. (2009).
Experimental test of quantum contextuality in neutron interferometry. Physical Review Letters,
103(4), 040403. https://doi.org/10.1103/PhysRevLett.103.040403, arXiv:0904.4576.
Belinfante, F. J. (1973). A survey of hidden-variables theories (International series of monographs
in natural philosophy, Vol. 55). Oxford/New York: Pergamon Press/Elsevier. https://doi.org/10.
1016/B978-0-08-017032-9.50001-7.
Bell, J. S. (1966). On the problem of hidden variables in quantum mechanics. Reviews of Modern
Physics, 38, 447–452. https://doi.org/10.1103/RevModPhys.38.447.
Bell, J. S. (1990). Against ‘measurement’. Physics World, 3, 33–41. https://doi.org/10.1088/2058-
7058/3/8/26.
Beltrametti, E. G., & Bugajski, S. (1996). The Bell phenomenon in classical frameworks. Journal
of Physics A: Mathematical and General Physics, 29. https://doi.org/10.1088/0305-4470/29/2/
005.
Beltrametti, E. G., & Maçzyński, M. J. (1991). On a characterization of classical and nonclassical
probabilities. Journal of Mathematical Physics, 32. https://doi.org/10.1063/1.529326.
Beltrametti, E. G., & Maçzyński, M. J. (1993). On the characterization of probabilities: A
generalization of Bell’s inequalities. Journal of Mathematical Physics, 34. https://doi.org/10.
1063/1.530333.
538 K. Svozil

Beltrametti, E. G., & Maçzyński, M. J. (1994). On Bell-type inequalities. Foundations of Physics,

24. https://doi.org/10.1007/bf02057861.
Beltrametti, E. G., & Maçzyński, M. J. (1995). On the range of non-classical probability. Reports
on Mathematical Physics, 36, https://doi.org/10.1016/0034-4877(96)83620-2.
Beltrametti, E. G., Del Noce, C, & Maçzyński, M. J. (1995). Characterization and deduction of
Bell-type inequalities. In: C. Garola & A. Rossi (Eds.), The foundations of quantum mechanics
– Historical analysis and open questions, Lecce, 1993 (pp. 35–41). Dordrecht: Springer. https://
doi.org/10.1007/978-94-011-00298_3.
Bishop, E., & Leeuw, K. D. (1959). The representations of linear functionals by measures on sets
of extreme points. Annals of the Fourier Institute, 9, 305–331. https://doi.org/10.5802/aif.95.
Bohr, N. (1949). Discussion with Einstein on epistemological problems in atomic physics. In P.
A. Schilpp (Ed.), Albert Einstein: Philosopher-Scientist, the library of living philosophers,
Evanston (pp. 200–241). https://doi.org/10.1016/S1876-0503(08)70379-7.
Boole, G. (1854). An investigation of the laws of thought. http://www.gutenberg.org/ebooks/
15114.
Boole, G. (1862). On the theory of probabilities. Philosophical Transactions of the Royal Society
of London, 152, 225–252. https://doi.org/10.1098/rstl.1862.0015.
Boschi, D., Branca, S., De Martini, F., & Hardy, L. (1997). Ladder proof of nonlocality without
inequalities: Theoretical and experimental results. Physical Review Letters, 79, 2755–2758.
https://doi.org/10.1103/PhysRevLett.79.2755.
Bub, J., & Demopoulos, W. (2010). Itamar Pitowsky 1950–2010. Studies in History and Philosophy
of Science Part B: Studies in History and Philosophy of Modern Physics, 41(2), 85. https://doi.
org/10.1016/j.shpsb.2010.03.004.
Bub, J., & Stairs, A. (2009). Contextuality and nonlocality in ‘no signaling’ theories. Foundations
of Physics, 39. https://doi.org/10.1007/s10701-009-9307-8.
Bub, J., & Stairs, A. (2010). Contextuality in quantum mechanics: Testing the Klyachko inequality.
https://arxiv.org/abs/1006.0500, arXiv:1006.0500.
Cabello, A. (1994). A simple proof of the Kochen-Specker theorem. European Journal of Physics,
15(4), 179–183. https://doi.org/10.1088/0143-0807/15/4/004.
Cabello, A. (1996). Pruebas algebraicas de imposibilidad de variables ocultas en mecánica
cuántica. PhD thesis, Universidad Complutense de Madrid, Madrid. http://eprints.ucm.es/1961/
1/T21049.pdf.
Cabello, A. (1997). No-hidden-variables proof for two spin- particles preselected and postselected
in unentangled states. Physical Review A, 55, 4109–4111. https://doi.org/10.1103/PhysRevA.
55.4109.
Cabello, A. (2008). Experimentally testable state-independent quantum contextuality. Phys-
ical Review Letters, 101(21), 210401. https://doi.org/10.1103/PhysRevLett.101.210401,
arXiv:0808.2456.
Cabello, A., & García-Alcaine, G. (1995). A hidden-variables versus quantum mechanics exper-
iment. Journal of Physics A: Mathematical and General Physics, 28. https://doi.org/10.1088/
0305-4470/28/13/016.
Cabello, A., Estebaranz, J. M., & García-Alcaine, G. (1996). Bell-Kochen-Specker theorem:
A proof with 18 vectors. Physics Letters A 212(4), 183–187. https://doi.org/10.1016/0375-
9601(96)00134-X, arXiv:quant-ph/9706009.
Cabello, A., Filipp, S., Rauch, H., & Hasegawa, Y. (2008). Proposed experiment for testing
quantum contextuality with neutrons. Physical Review Letters, 100, 130404. https://doi.org/
10.1103/PhysRevLett.100.130404.
Cabello, A., Severini, S., & Winter, A. (2010). (Non-)contextuality of physical theories as an
axiom. https://arxiv.org/abs/1010.2163, arXiv:1010.2163
Cabello, A., Badziag, P., Terra Cunha, M., & Bourennane, M. (2013). Simple Hardy-like proof
of quantum contextuality. Physical Review Letters, 111, 180404. https://doi.org/10.1103/
PhysRevLett.111.180404.
24 Roots and (Re)sources of Value (In)definiteness Versus Contextuality 539

Cabello, A., Severini, S., & Winter, A. (2014). Graph-theoretic approach to quantum correla-
tions. Physical Review Letters, 112, 040401. https://doi.org/10.1103/PhysRevLett.112.040401,
arXiv:1401.7081.
Cabello, A., Portillo, J. R., Solís, A., & Svozil, K. (2018). Minimal true-implies-false and true-
implies-true sets of propositions in noncontextual hidden-variable theories. Physical Review A,
98, 012106. https://doi.org/10.1103/PhysRevA.98.012106, arXiv:1805.00796.
Chen, J. L., Cabello, A., Xu, Z. P., Su, H. Y., Wu, C., & Kwek, L. C. (2013) Hardy’s paradox for
high-dimensional systems. Physical Review A, 88, 062116. https://doi.org/10.1103/PhysRevA.
88.062116.
Cirel’son (=Tsirel’son), B. S. (1980). Quantum generalizations of Bell’s inequality. Letters in
Mathematical Physics, 4(2), 93–100. https://doi.org/10.1007/BF00417500.
Cirel’son (=Tsirel’son), B. S. (1993). Some results and problems on quantum Bell-type inequali-
ties. Hadronic Journal Supplement, 8, 329–345. http://www.tau.ac.il/~tsirel/download/hadron.
pdf.
Clauser, J. (2002). Early history of Bell’s theorem. In R. Bertlmann & A. Zeilinger (Eds.), Quantum
(un)speakables: From Bell to quantum information (pp. 61–96). Berlin: Springer. https://doi.
org/10.1007/978-3-662-05032-3_6.
Clifton, R. K. (1993). Getting contextual and nonlocal elements-of-reality the easy way. American
Journal of Physics, 61(5), 443–447. https://doi.org/10.1119/1.17239.
Colins, D., & Gisin, N. (2004). A relevant two qbit Bell inequality inequivalent to the CHSH
inequality. Journal of Physics A: Mathematical and General, 37, 1775–1787. https://doi.org/
10.1088/0305-4470/37/5/021, arXiv:quant-ph/0306129.
Del Noce, C. (1995). An algorithm for finding Bell-type inequalities. Foundations of Physics
Letters, 8. https://doi.org/10.1007/bf02187346.
Dvurečenskij, A., & Länger, H. (1994). Bell-type inequalities in horizontal sums of boolean
algebras. Foundations of Physics, 24. https://doi.org/10.1007/bf02057864.
Dvurečenskij, A., & Länger, H. (1995a). Bell-type inequalities in orthomodular lattices. I.
Inequalities of order 2. International Journal of Theoretical Physics, 34. https://doi.org/10.
1007/bf00671363.
Dvurečenskij, A., & Länger, H. (1995b). Bell-type inequalities in orthomodular lattices. II.
Inequalities of higher order. International Journal of Theoretical Physics, 34. https://doi.org/
10.1007/bf00671364.
Dzhafarov, E. N., Cervantes, V. H., & Kujala, J. V. (2017). Contextuality in canonical systems
of random variables. Philosophical Transactions of the Royal Society A: Mathematical,
Physical and Engineering Sciences, 375(2106), 20160389. https://doi.org/10.1098/rsta.2016.
0389, arXiv:1703.01252.
Earman, J. (2007). Aspects of determinism in modern physics. Part B. In J. Butterfield & J.
Earman (Eds.), Philosophy of physics, handbook of the philosophy of science (pp. 1369–1434).
Amsterdam: North-Holland. https://doi.org/10.1016/B978-044451560-5/50017-8.
ESI – The Erwin Schrödinger International Institute for Mathematical Physics. (2001). Scientific
report for the year 2000. https://www.esi.ac.at/material/scientific-reports-1/2000.pdf, eSI-
Report 2000.
Everett III, H. (1973). The many-worlds interpretation of quantum mechanics (pp. 3–140).
Princeton: Princeton University Press.
Filipp, S., & Svozil, K. (2004a). Generalizing Tsirelson’s bound on Bell inequalities using a min-
max principle. Physical Review Letters, 93, 130407. https://doi.org/10.1103/PhysRevLett.93.
130407, arXiv:quant-ph/0403175.
Filipp, S., & Svozil, K. (2004b). Testing the bounds on quantum probabilities. Physical Review A,
69, 032101. https://doi.org/10.1103/PhysRevA.69.032101, arXiv:quant-ph/0306092.
540 K. Svozil

Fréchet, M. (1935). Généralisation du théorème des probabilités totales. Fundamenta Mathemati-

cae, 25(1), 379–387. http://eudml.org/doc/212798.
Froissart, M. (1981). Constructive generalization of Bell’s inequalities. Il Nuovo Cimento B (1971–
1996), 64, 241–251. https://doi.org/10.1007/BF02903286.
(aka George Orwell), E. A. B. (1949). Nineteen Eighty-Four (aka 1984). Twentieth century
classics, Secker & Warburg, Cambridge. http://gutenberg.net.au/ebooks01/0100021.txt.
Gleason, A. M. (1957). Measures on the closed subspaces of a Hilbert space. Journal of
Mathematics and Mechanics (now Indiana University Mathematics Journal), 6(4), 885–893.
https://doi.org/10.1512/iumj.1957.6.56050.
Godsil, C. D., & Zaks, J. (1988, 2012) Coloring the sphere, https://arxiv.org/abs/1201.0486,
University of Waterloo research report CORR 88-12, arXiv:1201.0486.
Grangier, P. (2002). Contextual objectivity: A realistic interpretation of quantum mechanics.
European Journal of Physics, 23(3), 331–337. https://doi.org/10.1088/0143-0807/23/3/312,
arXiv:quant-ph/0012122.
Grötschel, M., Lovász, L., & Schrijver, A. (1986). Relaxations of vertex packing. Journal of Com-
binatorial Theory, Series B, 40(3), 330–343. https://doi.org/10.1016/0095-8956(86)90087-0.
Hailperin, T. (1965) Best possible inequalities for the probability of a logical function of events.
The American Mathematical Monthly, 72(4), 343–359, https://doi.org/10.2307/2313491.
Hailperin, T. (1986). Boole’s logic and probability: Critical exposition from the standpoint of
contemporary algebra, logic and probability theory (Studies in logic and the foundations of
mathematics, Vol. 85, 2nd ed.). Elsevier Science Ltd. https://www.elsevier.com/books/booles-
logic-and-probability/hailperin/978-0-444-87952-3.
Hardy, L. (1992). Quantum mechanics, local realistic theories, and lorentz-invariant realistic
theories. Physical Review Letters, 68, 2981–2984. https://doi.org/10.1103/PhysRevLett.68.
2981.
Hardy, L. (1993). Nonlocality for two particles without inequalities for almost all entangled states.
Physical Review Letters, 71, 1665–1668. https://doi.org/10.1103/PhysRevLett.71.1665.
Hasegawa, Y., Loidl, R., Badurek, G., Baron, M., & Rauch, H. (2006). Quantum contextuality in a
single-neutron optical experiment. Physical Review Letters, 97(23), 230401. https://doi.org/10.
1103/PhysRevLett.97.230401.
Havlicek, H., Krenn, G., Summhammer, J., & Svozil, K. (2001). Colouring the rational quantum
sphere and the Kochen-Specker theorem. Journal of Physics A: Mathematical and General, 34,
3071–3077. https://doi.org/10.1088/0305-4470/34/14/312, arXiv:quant-ph/9911040.
Hrushovski, E., & Pitowsky, I. (2004). Generalizations of Kochen and Specker’s theorem and
the effectiveness of Gleason’s theorem. Studies in History and Philosophy of Science Part B:
Studies in History and Philosophy of Modern Physics, 35(2), 177–194. https://doi.org/10.1016/
j.shpsb.2003.10.002, arXiv:quant-ph/0307139
Johansen, H. B. (1994). Comment on “getting contextual and nonlocal elements-of-reality the easy
way”. American Journal of Physics, 62. https://doi.org/10.1119/1.17551.
Kleene, S. C. (1936). General recursive functions of natural numbers. Mathematische Annalen,
112(1), 727–742. https://doi.org/10.1007/BF01565439.
Klyachko, A. A. (2002). Coherent states, entanglement, and geometric invariant theory. https://
arxiv.org/abs/quant-ph/0206012, arXiv:quant-ph/0206012.
Klyachko, A. A., Can, M. A., Binicioğlu, S., & Shumovsky, A. S. (2008). Simple test for hidden
variables in spin-1 systems. Physical Review Letters, 101, 020403. https://doi.org/10.1103/
PhysRevLett.101.020403, arXiv:0706.0126
Kochen, S., & Specker, E. P. (1965). Logical structures arising in quantum theory. In The theory
of models, proceedings of the 1963 international symposium at Berkeley (pp. 177–189).
Amsterdam/New York/Oxford: North Holland. https://www.elsevier.com/books/the-theory-of-
models/addison/978-0-7204-2233-7, reprinted in Specker (1990, pp. 209–221).
Kochen, S., & Specker, E. P. (1967). The problem of hidden variables in quantum mechanics.
Journal of Mathematics and Mechanics (now Indiana University Mathematics Journal), 17(1),
59–87. https://doi.org/10.1512/iumj.1968.17.17004.
24 Roots and (Re)sources of Value (In)definiteness Versus Contextuality 541

Lakatos, I. (1978). Philosophical papers. 1. The methodology of scientific research programmes.

Cambridge: Cambridge University Press.
Länger, H., Maçzyński, M. J. (1995). On a characterization of probability measures on boolean
algebras and some orthomodular lattices. Mathematica Slovaca, 45(5), 455–468. http://eudml.
org/doc/32311.
Lovász, L., Saks, M., & Schrijver, A. (1989). Orthogonal representations and connectivity of
graphs. Linear Algebra and Its Applications, 114–115, 439–454. https://doi.org/10.1016/0024-
3795(89)90475-8, special Issue Dedicated to Alan J. Hoffman.
Marcus, M., & Ree, R. (1959). Diagonals of doubly stochastic matrices. The Quarterly Journal of
Mathematics, 10(1), 296–302. https://doi.org/10.1093/qmath/10.1.296.
Mermin, D. N. (1989a). Could Feynman have said this? Physics Today, 57, 10–11. https://doi.org/
10.1063/1.1768652.
Mermin, D. N. (1989b) What’s wrong with this pillow? Physics Today, 42, 9–11. https://doi.org/
10.1063/1.2810963.
Mermin, D. N. (2007). Quantum computer science. Cambridge: Cambridge University Press.
https://doi.org/10.1017/CBO9780511813870.
Meyer, D. A. (1999). Finite precision measurement nullifies the Kochen-Specker theorem.
Physical Review Letters, 83(19), 3751–3754. https://doi.org/10.1103/PhysRevLett.83.3751.
arXiv:quant-ph/9905080.
Moore, E. F. (1956). Gedanken-experiments on sequential machines. In C. E. Shannon, & J.
McCarthy (Eds.), Automata studies, (AM-34) (pp. 129–153). Princeton: Princeton University
Press. https://doi.org/10.1515/9781400882618-006.
Nietzsche, F. (1887, 2009–). Zur Genealogie der Moral (On the Genealogy of Morality). http://
www.nietzschesource.org/#eKGWB/GM, digital critical edition of the complete works and
letters, based on the critical text by G. Colli & M. Montinari, Berlin/New York, de Gruyter
1967-, edited by Paolo D’Iorio.
Nietzsche, F. W. (1887, 1908; 1989, 2010). On the genealogy of morals and Ecce homo. Vintage,
Penguin, Random House. https://www.penguinrandomhouse.com/books/121939/on-the-
genealogy-of-morals-and-ecce-homo-by-friedrich-nietzsche-edited-with-a-commentary-by-
walter-kaufmann/9780679724629/, translated by Walter Arnold Kaufmann.
Odifreddi, P. (1989). Classical recursion theory (Vol. 1). Amsterdam: North-Holland.
Parsons, T., & Pisanski, T. (1989) Vector representations of graphs. Discrete Mathematics, 78,
https://doi.org/10.1016/0012-365x(89)90171-4.
Peres, A. (1978). Unperformed experiments have no results. American Journal of Physics, 46,
745–747, https://doi.org/10.1119/1.11393.
Peres, A. (1993). Quantum theory: Concepts and methods. Dordrecht: Kluwer Academic Publish-
ers.
Pitowsky, I. (1982). Substitution and truth in quantum logic. Philosophy of Science, 49. https://doi.
org/10.2307/187281.
Pitowsky, I. (1983). Deterministic model of spin and statistics. Physical Review D, 27, 2316–2326.
https://doi.org/10.1103/PhysRevD.27.2316.
Pitowsky, I. (1986). The range of quantum probabilities. Journal of Mathematical Physics, 27(6),
1556–1565.
Pitowsky, I. (1989a). From George Boole to John Bell: The origin of Bell’s inequality. In M.
Kafatos (Ed.), Bell’s theorem, quantum theory and the conceptions of the universe (Fundamen-
tal theories of physics, Vol. 37, pp. 37–49). Dordrecht: Kluwer Academic Publishers/Springer.
https://doi.org/10.1007/978-94-017-0849-4_6.
Pitowsky, I. (1989b). Quantum probability – Quantum logic (Lecture notes in physics, Vol. 321).
Berlin/Heidelberg: Springer. https://doi.org/10.1007/BFb0021186.
Pitowsky, I. (1991). Correlation polytopes their geometry and complexity. Mathematical Program-
ming, 50, 395–414. https://doi.org/10.1007/BF01594946.
Pitowsky, I. (1994). George Boole’s ‘conditions of possible experience’ and the quantum puzzle.
The British Journal for the Philosophy of Science, 45, 95–125. https://doi.org/10.1093/bjps/45.
1.95.
542 K. Svozil

Pitowsky, I. (1998). Infinite and finite Gleason’s theorems and the logic of indeterminacy. Journal
of Mathematical Physics, 39(1), 218–228. https://doi.org/10.1063/1.532334.
Pitowsky, I. (2003). Betting on the outcomes of measurements: A Bayesian theory of quantum
probability. Studies in History and Philosophy of Science Part B: Studies in History and Phi-
losophy of Modern Physics, 34(3), 395–414. https://doi.org/10.1016/S1355-2198(03)00035-2,
quantum Information and Computation, arXiv:quant-ph/0208121.
Pitowsky, I. (2006). Quantum mechanics as a theory of probability. In W. Demopoulos &
I. Pitowsky (Eds.), Physical theory and its interpretation (The western Ontario series in
philosophy of science, Vol. 72, pp. 213–240). Springer Netherlands. https://doi.org/10.1007/
1-4020-4876-9_10, arXiv:quant-ph/0510095.
Pitowsky, I., & Svozil, K. (2001). New optimal tests of quantum nonlocality. Physical Review A,
64, 014102. https://doi.org/10.1103/PhysRevA.64.014102, arXiv:quant-ph/0011060.
Planck, M. (1932). The concept of causality. Proceedings of the Physical Society, 44(5), 529–539.
https://doi.org/10.1088/0959-5309/44/5/301.
Pták, P., & Pulmannová, S. (1991). Orthomodular structures as quantum logics. Intrinsic properties,
state space and probabilistic topics (Fundamental theories of physics, Vol. 44). Dordrecht:
Kluwer Academic Publishers/Springer.
Pulmannová, S. (2002). Hidden variables and Bell inequalities on quantum logics. Foundations of
Physics, 32, https://doi.org/10.1023/a:1014424425657.
Pykacz, J., & Santos, E. (1991). Hidden variables in quantum logic approach reexamined. Journal
of Mathematical Physics, 32. https://doi.org/10.1063/1.529327.
Ramanathan, R., Rosicka, M., Horodecki, K., Pironio, S., Horodecki, M., & Horodecki, P. (2018).
Gadget structures in proofs of the Kochen-Specker theorem. https://arxiv.org/abs/1807.00113,
arXiv:1807.00113
Rogers, Jr H. (1967). Theory of recursive functions and effective computability. New
York/Cambridge, MA: MacGraw-Hill/The MIT Press.
Schaller, M., & Svozil, K. (1995). Automaton partition logic versus quantum logic. International
Journal of Theoretical Physics, 34(8), 1741–1750. https://doi.org/10.1007/BF00676288.
Schaller, M., & Svozil, K. (1996). Automaton logic. International Journal of Theoretical Physics,
35, https://doi.org/10.1007/BF02302381.
Segal, A., & Goldschmidt, T. (2017/2018). The necessity of idealism. In Idealism: New essays
in metaphysics (pp. 34–49). Oxford: Oxford University Press. https://doi.org/10.1093/oso/
9780198746973.003.0003.
Simmons, A. W. (2020). How (maximally) contextual is quantum mechanics? This volume,
pp. 505–520.
Sliwa, C. (2003). Symmetries of the Bell correlation inequalities. Physics Letters A, 317, 165–168.
https://doi.org/10.1016/S0375-9601(03)01115-0, arXiv:quant-ph/0305190.
Smullyan, R. M. (1993). Recursion theory for metamathematics (Oxford logic guides, Vol. 22).
New York/Oxford: Oxford University Press.
Solís-Encina, A., & Portillo, J. R. (2015). Orthogonal representation of graphs. https://arxiv.org/
abs/1504.03662, arXiv:1504.03662.
Specker, E. (1960). Die Logik nicht gleichzeitig entscheidbarer Aussagen. Dialectica, 14(2–3),
239–246. https://doi.org/10.1111/j.1746-8361.1960.tb00422.x, arXiv:1103.4537.
Specker, E. (1990). Selecta. Basel: Birkhäuser Verlag. https://doi.org/10.1007/978-3-0348-9259-
9.
Specker, E. (2009). Ernst Specker and the fundamental theorem of quantum mechanics. https://
vimeo.com/52923835, video by Adán Cabello, recorded on June 17, 2009.
Stace, W. T. (1934). The refutation of realism. Mind, 43(170), 145–155. https://doi.org/10.1093/
mind/XLIII.170.145.
Stairs, A. (1983). Quantum logic, realism, and value definiteness. Philosophy of Science, 50, 578–
602. https://doi.org/10.1086/289140.
Svozil, K. (2001). On generalized probabilities: Correlation polytopes for automaton logic and
generalized urn models, extensions of quantum mechanics and parameter cheats. https://arxiv.
org/abs/quant-ph/0012066, arXiv:quant-ph/0012066.
24 Roots and (Re)sources of Value (In)definiteness Versus Contextuality 543

Svozil, K. (2002). Quantum information in base n defined by state partitions. Physical Review A,
66, 044306. https://doi.org/10.1103/PhysRevA.66.044306, arXiv:quant-ph/0205031.
Svozil, K. (2004). Quantum information via state partitions and the context translation principle.
Journal of Modern Optics, 51, 811–819. https://doi.org/10.1080/09500340410001664179,
arXiv:quant-ph/0308110.
Svozil, K. (2005). Logical equivalence between generalized urn models and finite automata.
International Journal of Theoretical Physics, 44, 745–754. https://doi.org/10.1007/s10773-
005-7052-0, arXiv:quant-ph/0209136.
Svozil, K. (2009a). Proposed direct test of a certain type of noncontextuality in quantum
mechanics. Physical Review A, 80(4), 040102. https://doi.org/10.1103/PhysRevA.80.040102.
Svozil, K. (2009b). Quantum scholasticism: On quantum contexts, counterfactuals, and the
absurdities of quantum omniscience. Information Sciences, 179, 535–541. https://doi.org/10.
1016/j.ins.2008.06.012.
Svozil, K. (2012). How much contextuality? Natural Computing, 11(2), 261–265. https://doi.org/
10.1007/s11047-012-9318-9, arXiv:1103.3980.
Svozil, K. (2016). Quantum hocus-pocus. Ethics in Science and Environmental Politics (ESEP),
16(1), 25–30. https://doi.org/10.3354/esep00171, arXiv:1605.08569.
Svozil, K. (2017a). Classical versus quantum probabilities and correlations. https://arxiv.org/abs/
1707.08915, arXiv:1707.08915.
Svozil, K. (2017b). Quantum clouds. https://arxiv.org/abs/1808.00813, arXiv:1808.00813.
Svozil, K. (2018a). Kolmogorov-type conditional probabilities among distinct contexts. https://
arxiv.org/abs/1903.10424, arXiv:1903.10424.
Svozil, K. (2018b). New forms of quantum value indefiniteness suggest that incompatible
views on contexts are epistemic. Entropy, 20(6), 406(22). https://doi.org/10.3390/e20060406,
arXiv:1804.10030.
Svozil, K. (2018c). Physical [a]causality. Determinism, randomness and uncaused events.
Cham/Berlin/Heidelberg/New York: Springer. https://doi.org/10.1007/978-3-319-70815-7.
Sylvia, P., & Majernik, V. (1992). Bell inequalities on quantum logics. Journal of Mathematical
Physics, 33. https://doi.org/10.1063/1.529638.
Ursic, S. (1984). A linear characterization of NP-complete problems. In R. E. Shostak (Ed.), 7th
international conference on automated deduction, Napa, 14–16 May 1984 Proceedings (pp. 80–
100). New York: Springer. https://doi.org/10.1007/978-0-387-34768-4_5.
Ursic, S. (1986). Generalizing fuzzy logic probabilistic inferences. In Proceedings of the second
conference on uncertainty in artificial intelligence, UAI’86 (pp. 303–310). Arlington: AUAI
Press. http://dl.acm.org/citation.cfm?id=3023712.3023752, arXiv:1304.3114.
Ursic, S. (1988). Generalizing fuzzy logic probabilistic inferences. In J. F. Lemmer & L. N. Kanal
(Eds.), Uncertainty in artificial intelligence 2 (UAI1986) (pp. 337–362), Amsterdam: North
Holland.
Vermaas, P. E. (1994). Comment on “getting contextual and nonlocal elements-of-reality the easy
way”. American Journal of Physics, 62. https://doi.org/10.1119/1.17488.
Vorob’ev, N. N. (1962). Consistent families of measures and their extensions. Theory of Probability
and Its Applications, 7. https://doi.org/10.1137/1107014.
Weihs, G., Jennewein, T., Simon, C., Weinfurter, H., & Zeilinger, A. (1998). Violation of Bell’s
inequality under strict Einstein locality conditions. Physical Review Letters, 81, 5039–5043.
https://doi.org/10.1103/PhysRevLett.81.5039.
Wright, R. (1978). The state of the pentagon. A nonclassical example. In A. R. Mar-
low (Ed.), Mathematical foundations of quantum theory (pp. 255–274). New York: Aca-
demic Press. https://www.elsevier.com/books/mathematical-foundations-of-quantum-theory/
marlow/978-0-12-473250-6.
Wright, R. (1990). Generalized urn models. Foundations of Physics, 20(7), 881–903. https://doi.
org/10.1007/BF01889696.
544 K. Svozil

Zeilinger, A. (1999). A foundational principle for quantum mechanics. Foundations of Physics,

29(4), 631–643. https://doi.org/10.1023/A:1018820410908.
Zierler, N., & Schlessinger, M. (1965). Boolean embeddings of orthomodular sets and quantum
logic. Duke Mathematical Journal, 32, 251–262. https://doi.org/10.1215/S0012-7094-65-
03224-2, reprinted in Zierler and Schlessinger (1975).
Zierler, N., & Schlessinger, M. (1975). Boolean embeddings of orthomodular sets and quantum
logic. In C. A. Hooker (Ed.), The logico-algebraic approach to quantum mechanics: Volume I:
Historical evolution (pp. 247–262). Dordrecht: Springer Netherlands. https://doi.org/10.1007/
978-94-010-1795-4_14.
Chapter 25
Schrödinger’s Reaction to the EPR Paper

Jos Uffink

Abstract I discuss Schrödinger’s response to the Einstein-Podolsky-Rosen (EPR)

paper of 1935. In particular, it is argued, based on an unpublished notebook, that
while Schrödinger sympathized with the EPR argument he worried about its lack
of emphasis on the role of biorthogonal expansions and on the distinction between
conclusions dependent on the particular outcome obtained in a single measurement,
versus those that quantify over all possible outcomes of a measurement procedure.
I also discuss the different views between Schrödinger and Einstein on the issue
of the completeness of quantum mechanics. Finally, I discuss the question whether
Schrödinger borrowed from his correspondence with Einstein to come up with his
famous Cat Paradox.

Keywords Schrödinger · Quantum mechanics · Einstein-Podolski-Rosen

argument · Completeness · Separability · Schrödinger’s cat

25.1 Introduction

On March 25, 1935, the Physical Review received a paper by Einstein, Podolsky
and Rosen (EPR), arguing that the quantum description of physical reality was
incomplete. It was published on May 15 of the same year. It will be superfluous, on
this occasion, to stress the importance of this paper for the subsequent discussions
of the foundations of quantum theory in the 1930s and beyond, in particular
by directing these discussions to focus on the phenomenon of entanglement,
eventually opening up, in recent years, investigations into quantum information
and computation. It is a pleasure to dedicate this paper to the memory of Itamar

J. Uffink ()
Philosophy Department and Program in History of Science, Technology and Medicine,
University of Minnesota, Minneapolis, MN, USA
e-mail: jbuffink@umn.edu

© Springer Nature Switzerland AG 2020 545

M. Hemmo, O. Shenker (eds.), Quantum, Probability, Logic,
Jerusalem Studies in Philosophy and History of Science,
https://doi.org/10.1007/978-3-030-34316-3_25
546 J. Uffink

Pitowsky, who did so much conceptual work to understand quantum entanglement

and to distinguish it from mere classical correlations.
Actually, the term ‘entanglement’ does not appear in the EPR paper. This
term (or its German equivalent ‘Verschränkung’) was coined by Schrödinger in
two papers he wrote during the following summer (1935a, 1935b) and finished
nearly simultaneously around 11 August. One paper, for a general audience,
was published in Die Naturwissenschaften and the second, more technical, in
the Mathematical Proceedings of the Cambridge Philosophical Society, where he
described entanglement as “not one but the characteristic trait of quantum mechanics
that enforces its entire departure from classical lines of thought.”
It is commonly assumed in the historical literature that Schrödinger’s papers were
triggered by the EPR paper, and indeed, Schrödinger acknowledged as much in a
footnote in (1935b, p. 845), where, referring to the EPR paper, he wrote
The appearance of this work gave the incentive for the present — shall I say seminar paper
or general report?

However, although EPR obviously stimulated Schrödinger to publish his thoughts,

this is not to say that he developed these thoughts only after he had learned
of the EPR paper. Elsewhere, Christoph Lehner and I will present evidence
from Schrödinger’s unpublished notebooks, and (partly published) letters that
detail Schrödinger’s role in the development of the EPR argument. In particular,
these notebooks and letters show how Schrödinger struggled with the concept of
entanglement already from 1926-7 onwards, and how he essentially developed what
we now call the EPR argument already in November 1931, 3 12 years before EPR,
when he and Einstein were still colleagues in Berlin. Here, however, I will focus
on Schrödinger’s immediate reactions to the appearance of the EPR paper, and his
subsequent correspondence with Einstein in 1935. Schrödinger wrote to Einstein on
June 7, 1935, to express his joy at the appearance of the EPR paper, “about what
we discussed so much in Berlin”, probably referring to discussions in November
1931. However, Schrödinger’s notebook entitled Übungen1 contains notes which
he took while preparing this letter. These notes clearly reveal that Schrödinger was
not entirely happy with the presentation of the argument in the EPR paper. He also
expressed his worries in the letter to Einstein:
Can I add a few words? They may look at first sight like objections, but they are just points
that I would like to have formulated more clearly (von Meyenn 2011, p. 527).

The points that Schrödinger worried about are subtle. They concern both the role of
biorthogonal decompositions of bipartite wave functions and the crucial dependence
on the measurement outcomes in the EPR argument. To understand his worries, it
might be well to first recall the structure of the argument in the EPR paper.

1 Thisnotebook is accessible at https://fedora.phaidra.univie.ac.at/fedora/objects/o:159711/

methods/bdef:Book/view. The relevant text is on pp. 35–41.
25 Schrödinger’s Reaction to the EPR Paper 547

25.2 The Argumentation of the EPR Paper

EPR present a necessary Criterion for Completeness of a physical theory,2 and a

sufficient Criterion for Reality for what counts as “an element of physical reality”.3
After rehearsing some well-known aspects of quantum mechanics in their section 1,
their section 2 introduces the example of two systems that have interacted in some
period in the past, but between which any interaction has ceased since then.4
EPR then present a wave function for such a joint system, which can be expressed
in different forms5 :
 
|
 = cn |an 1 |ψ n 2 = cn |an 1 |φ n 2 (25.1)
n n

Here, {|an } denotes an orthonormal basis of eigenvectors of some physical quantity

A of system 1 and {|an } another orthonormal basis of eigenvectors of quantity
A (which need not commute with A) of system 1. {|ψ n } and {|φ n } denote two
arbitrary collections of (linearly independent) unit vectors for system 2, neither of
which needs to be mutually orthogonal nor complete. Note that the expansions of
the wave function in (25.1) are not biorthogonal.
EPR argue that if we make a measurement of quantity A on system 1, say with
result ak , the second system is to be assigned a wave function corresponding to
the vector |ψ k , by the process of “the reduction of the wave packet” (or in more
modern terms: by applying the projection postulate). However, if we decide instead
to make a measurement of quantity A on system 1, and obtain an eigenvalue ar ,
the corresponding wave function for system 2 is |φ r . But, as EPR argue, what is
physically real for system 2 should not depend on what is done to system 1, since
the two systems no longer interact. Thus, they draw the following, what I will refer
to as their ‘general conclusion’ from their argument:

2 “[E]very element of the physical reality must have a counterpart in the physical theory.” (Einstein
et al. 1935, p. 777)
3 “If, without in any way disturbing a system, we can predict with certainty (i.e., with probability

equal to unity) the value of a physical quantity, then there exists an element of reality corresponding
to that quantity.” (Einstein et al. 1935, p. 777)
4 Incidentally, the EPR paper never refers to spatial separation between the two systems as a

condition that would guarantee the absence of such interaction, unlike Schrödinger’s argument
in 1931.
5 Here and in further quotations, I have not followed original notation. In particular, wave functions

are presented here in the now common Dirac notation for perspicuity, even though, strictly
speaking, symbols like |
 are vectors in a Hilbert space, and the corresponding wave function

(x1 , x2 ) =  x1 , x2 |
 is a representation of this vector in the position eigenbasis. Also, I abstain
from refering to such entities as ‘states’, because it is an issue of contention between Einstein and
Schrödinger whether wave functions represent states of a system.
548 J. Uffink

[GENERAL CONCLUSION:]“Thus, it is possible to assign two different wave functions (in

our example |ψ k  and |φ r ) to the same reality (the second system after its interaction with
the first).” (original italicization)

After reaching this general conclusion, EPR proceed to discuss a special case, in
which it happens that ψ k and φ r above are eigenfunctions of the non-commuting
quantities P and Q, respectively, of system 2. For this purpose, they employ the
wave function
 +∞
1

(q1 , q2 ) = √ ei(q1 −q2 +q0 )p dp (25.2)
2π −∞

which, in Dirac notation, corresponds to

 
1 1
|
 = √ e iq0 p
|p1 | − p2 dp = √ |q1 |q − q0 2 dq (25.3)
2π 2π

where q0 is a real-valued constant.

In this special case, measuring the momentum P1 on system 1, with an outcome
p, say, leads one to assign the wave function |−p2 to system 2, i.e. an eigenfunction
of the momentum P2 of system 2 with eigenvalue −p. Hence, we can infer the
prediction that, with 100% certainty, any measurement of the momentum of 2 will
provide an outcome −p, and this value, by the Criterion of Reality, should therefore
correspond to an element of physical reality. Similarly, if we were to measure the
position Q1 on system 1, and obtain an outcome q1 , we can not only assign the the
wave function |q1 − q0 2 to system 2 but also infer that any measurement of the
position Q2 on system 2 will 100% certainly lead to the value q1 − q0 . Hence by the
Criterion of Reality, in this case the position of system 2 corresponds to an element
of reality.
The conclusion from this special case can therefore be formulated in terms of
different physical quantities of system 2, depending on what we measure on system
1, instead of just different wave functions as in the general conclusion. Yet, just like
that previous conclusion, it is argued that the elements of reality for system 2 should
not depend on what is done to system 1 (i.e.: the choice of measuring either P1 or
Q1 ), so that the quantities for system 2 must represent simultaneous elements of
reality:
[SPECIAL CONCLUSION:]“ [. . . ] we arrived at the conclusion that two physical quantities,
with non-commuting operators, can have simultaneous reality.”

And this conclusion is then shown to lead to a contradiction with the assumption
that the wave function gives a complete description of reality, because quantum
mechanics does not allow any wave function for system 2 in which both P2 and Q2
are both elements of reality.
It is worth emphasizing that the argument for the special conclusion, unlike the
general conclusion, makes essential use of the circumstance that the special wave
function (25.3) possesses multiple biorthogonal decompositions.
25 Schrödinger’s Reaction to the EPR Paper 549

25.3 Schrödinger’s Objections to the EPR Argument

The notes in the undated folder Übungen relating to the EPR paper start off with
2 pages entitled “Entwicklung e. Funktion zweier Variablen in eine biorthogonale
Reihe” (Expansion of a function of two variables in a biorthogonal series). This is a
rehearsal of a proof of the now well-known biorthogonal decomposition theorem,—
a proof Schrödinger had in fact already worked out earlier, in the folder Supraleitung
1931.6 He included a proof of the theorem along the lines of these notes in his June 7
letter to Einstein, and also in Schrödinger (1935a). In Übungen, he did not work out
all the details of the theorem, but only to the point that he refreshed his knowledge
of the topic enough to continue. He writes:
The main issue established. To be worked out later.

Schrödinger then immediately turns to the EPR paper (ignoring Einstein’s coau-
thors) with the remark:
Einstein does not write of a complete biorthogonal decomposition.

And his argument ends with the conclusion:

Thus, I have to assume a biorthogonal decomposition.

Even before we go into the details of his argument, these remarks show two
points: he is contemplating what I called the general conclusion of the EPR
paper, based on the non-biorthogonal decompositions (25.1), rather than the special
conclusion, where (25.3) of course does provide biorthogonal decompositions.
Secondly: Schrödinger thought that this general conclusion was problematic pre-
cisely because it did not assume a biorthogonal decomposition. In more detail, his
argument in Übungen initially worries about the question what the non-biorthogonal
decompositions in (25.1) imply for the application of the projection postulate after
a measurement on system 1. But these initial worries do not materialize, the text
breaks off, and he then switches to a different problem:
However, the problem is different. From the mere fact that after one measurement [i.e. of A
on system 1] we are stuck with one [wave function ψ k ], but after another measurement [of
A on system 1] we are stuck with another [wave] function [φ r ], where these functions
correspond to contradictory properties of “the other” system [2], one cannot fabricate
a noose by which to hang quantum mechanics (läßt sich die Qu. mech. noch kein
Strick drehen). Because even one and the same method of measurement can and must, for
identically prepared systems, provide sometimes this, sometimes another outcome, which
lead to inferences about the “other” system that will likewise be contradictory.

Clearly, Schrödinger sees the general conclusion in the EPR paper as insufficient to
produce a paradox. He continues:
It is therefore also not a contradiction when the function [
(q1 , q2 )] includes the possibility
that on one occasion a measurement delivers a determinate value of position, and on another

6 https://fedora.phaidra.univie.ac.at/fedora/objects/o:259940/methods/bdef:Book/view, pp.95–6.
550 J. Uffink

occasion another measurement a determinate value of momentum. Because, indeed, it could

be that really in one case the position, in the other case the momentum of the “other” system
was sharp.
The contradiction only arises when I recognize the following:
1. There is one measurement arrangement that always lets me infer a determinate position
for the other system. And
2. There is another measurement arrangement that at least sometimes provides a value for
a quantity of the other system which is incompatible with determinateness of position.

Thus, I have to assume a biorthogonal decomposition:



(q1 , q2 ) = ci gi (q1 )ui (q2 ) (25.4)

What was he thinking here? Let me try to reconstruct the argument by breaking
the problem down in a piecemeal manner, proceeding from simple cases to more
general ones. There are two main themes throughout this reconstruction attempt: the
first is the distinction between biorthogonal and non-biorthogonal decompositions
of the joint wave function |
, as already indicated. The second theme is the
distinction between inferences based on just a single performance of a measurement
(Einzelmessung), which will naturally be conditional on the particular outcome
obtained in this single measurement, versus conclusions that can be stated by
quantifying over all possible outcomes of the measurement procedure, and therefore
are not conditional on any particular outcome.
The simplest case is when the state vector |
 for the pair of systems has a unique
biorthogonal decomposition, say:

|
 = ci |ai |bi  (25.5)
i

Indeed, according to the Biorthogonal Decomposition Theorem, this is the generic

case for almost all |
 ∈ H1 ⊗ H2 , i.e., as long as the absolute values of all
coefficients ci are different. If we measure quantity A on system 1, for this state, and
obtain the outcome ak , we assign wave function |bk  to system 2. But if we obtain a
different outcome, ar , one should assign the wave function |br  to system 2. Thus,
depending on which outcome we obtain, we should assign different wave functions
to system 2. But could we argue in this case, that because system 2 does not interact
with system 1, these different wave functions refer to the same physical reality for
system 2? Surely not, because because even if one agrees to EPR’s claim that “no
real change can take place in the second system in consequence of anything that may
be done to the first system”, in the case at hand we are not doing different things
on the first system at all (we measure the same quantity A in both cases, i.e. we are
executing “one and the same method of measurement” as Schrödinger writes in the
quote above.). Rather, the question is here whether the physical reality of system 2
should be independent of the outcome obtained in this measurement of A on system
25 Schrödinger’s Reaction to the EPR Paper 551

1.7 And the answer to that question is generally negative, even in classical physics,
because two systems, even when they do not interact, might be correlated, so that
the outcome obtained by a measurement on 1 can still be informative about what is
real for system 2. Indeed, one of the main themes that surface again and again in
Schrödinger’s writings on entanglement and the EPR problem is that they express
an intermixture of ‘mental’ (where today we would perhaps say ‘epistemic’) and
‘physical’ (where today we would perhaps say ‘ontic’) aspects of the wave function.
Therefore, in the special case of (25.5), the fact that, after the same measurement
yields different outcomes on system 1, different wave functions are assigned to
system 2, depending on the outcomes obtained on system 1, does not justify EPR’s
general conclusion that those wave functions represent the same physical reality for
system 2.
Let us proceed to a more general case, where the decomposition is not biorthog-
onal, i.e.:

|
 = cn |an |ψ n  (25.6)
n

where, as before, |ψ n  denote linearly independent wave functions for system 2,

not necessarily orthogonal or complete. However, for a given |
, and a given
orthonormal basis {|an }, this set is uniquely determined (up to inessential phase
factors). In fact, the set {|ψ n } forms what Everett in 1957 called the relative states
of system 2, relative to the choice of |
 and the choice of basis {|an } on H1 .
The main difference with the previous case (25.5) is just that now the inference
to “elements of physical reality” is not so simple. Wave functions are not the kind
of entities that can be predicted with 100% certainty, especially not when one has to
decide between a non-orthogonal set. However, this problem is alleviated by noting
that every wave function, say |ψ i , is an eigenfunction of some observable (or set
of observables). In particular, one could consider the projectors P i := |ψ i ψ i |
as physical quantities, each of which will have the property that when the wave
function of system 2 is |ψ i , one can predict with 100% probability that the physical
quantity P i takes the value 1. So, here too, one can argue that, conditional on a
measurement of quantity A on system 1 yielding an outcome ak , there is a quantity
P k of the second system that will have the value 1, even without actually intervening
on system 2 and therefore, by the EPR reality criterion, to the inference that the
quantity P k must correspond to an element of reality for system 2.
However, in this case too, we are still left with the fact that this inference is
conditional on the outcome obtained in the measurement on system 1. And given
that the particular outcome obtained in such a measurement might be informative
about what is real about system 2, in the simple case (25.5), it would be hard to
deny a similar possibility too in the case of (25.6): different possible outcomes of
the A-measurement on system 1, say ak and ar (k = r), are mutually exclusive data

7 Thispoint is reminiscent of the distinction between parameter independence and outcome

independence in later discussions of the Bell inequalities, cf. Jarrett, Shimony.
552 J. Uffink

obtained from system 1. If these outcomes are held to be informative about system
2, they might represent mutually exclusive pieces of information about system 2.
But then, how could one argue that the inferences about system 2, i.e. that either
of the quantities P k or P r represents an element of reality by having the value 1,
refer to the same reality? Accordingly, Schrödinger does not regard it as paradoxical
by itself that, conditional on different outcomes obtained in measurements on the
first system, one should be able to predict (and infer the reality of) values of non-
commuting quantities of the second system.
He expresses this worry in his June 7 letter to Einstein:
In order to construct a contradiction, it is in my view not sufficient that for the same
preparation of the system pair the following can happen: one concrete single measurement
on the first system leads to a determinate value of a quantity [B] of the second system, any
other value of [B] or [B  ] is excluded for general reasons. I believe this is not sufficient.
Indeed, the fact that same preparations do not always give the same measurement outcomes
is something we have to concede anyway. So when on one occasion the measurement gives
the value [bk ], and on another occasion the value [br ] to quantity B, why would it not give
on a third time, not the quantity B but the quantity B  reality, and provide it with a value
[bs ]?

The point that Schrödinger raises in the the last sentence, namely that the
measurement on system 1 could lead to different (in particular non-commuting)
quantities obtaining values, would be hard to appreciate without the context of his
notes in the folder Übungen, because Schrödinger does not explicitly mention the
issue of biorthogonality in this passage. But one can infer from these notes that
the non-biorthogonality of the EPR wave function (25.1) was his main concern,
and since my reconstruction argues that for a non-biorthogonal decomposition,
like (25.6), one would indeed end up assigning reality to values of non-commuting
quantities like P k and P r , depending on the outcome obtained in the measurement
on system 1, this last sentence just quoted makes sense.
But of course, the two cases just reviewed still do not cover the case treated by
EPR’s general conclusion, i.e. the state (25.1). Here, we consider a measurement
of A, or a measurement of B, on system 1 with outcome ak or br respectively, and
contemplate their implications for system 2. Both EPR and Schrödinger agree that
this choice of what is done to system 1 should not affect what is real on system 2.
(As he said in 1931: what we choose to measure depends on an arbitrary choice
(Willesentscheidung).) But what we infer about system 2 depends not only on what
we do to system 1, it depends also on the outcome we obtain. Now, it would be
hard to deny that, just as a1 and a2 are mutually incompatible outcomes for the
measurement of A on system 1, that allow different inferences about what is real
for system 2, in example (25.6), the outcomes ak and br obtained in incompatible
measurements on system 1 also provide incompatible data that enable different
inferences about what is real for system 2. And so one may well ask whether this
should be in any way more paradoxical than the case already discussed under (25.6).
Schrödinger’s answer seems to be negative.
To summarize, what I have been arguing so far is that Schrödinger’s misgivings
about the EPR argument focus on their general conclusion, and particularly on the
fact that, at this stage of their argument, biorthogonality of the decomposition of the
25 Schrödinger’s Reaction to the EPR Paper 553

wave function is not assumed. He argues that any inference about the second system,
based on a measurement of the first, depends not only on the quantity measured on
the first system, but also on the outcome obtained in such a measurement. In the
case of a non-biorthogonal decomposition of the wave function, he does not regard
it as paradoxical that, depending on the outcome of a measurement on system 1,
we should infer that the values of one or another of non-commuting quantities for
system 2 constitute an element of reality.
Schrödinger also describes in his June 7 letter to Einstein what he thinks is a
satisfactory way to argue for a contradiction. The argument is very similar to what
he wrote in his Übungen notebook:
In order to construct a contradiction, the following seems necessary to me. There should
be a wave function for the pair of systems and there should be two quantities A and [A ],
whose simultaneous reality is excluded for general reasons, and for those, the following
should hold:
1. There exists a method of measuring, which for this wave function always gives the
quantity A a sharp value (even if not always the same value), so that I can say, even
without actually performing the measurement: in the case of such a wave function, A
possesses reality — I don’t care about which value it has. (welchen Wert es hat ist mir
Wurst.)
2. Another method of measurement must provide, at least sometimes, a sharp value for the
quantity [A ] (always for the same wave function, of course).

Schrödinger then points out that the maximally entangled wave function used in
the example by EPR in their special conclusion does indeed satisfy these conditions,
(and in fact always provides a sharp value in both measurements), but that this
is due to the fact that their example wave function allows multiple biorthogonal
decompositions. In fact, this example wave function is exceptionally special in
the sense that the absolute values of the coefficients (eipx0 ) are all the same, and
this implies that by performing an arbitrary unitary basis transformation on both
bases in the Hilbert spaces of the two systems, one can get another biorthogonal
decomposition. He then expresses a secondary worry that a conclusion as far-
reaching as EPR, should not depend on such an exceptional example. But he also
points out that if one does not share this secondary worry, the example is just fine
in demonstrating the argument. His letter ends with an outline of a proof of the
biorthogonal decomposition theorem.
In this attempt at reconstructing Schrödinger’s thinking, the theme of the dis-
tinction between conclusions based single measurements versus conclusions based
on general conclusions that are not conditional on a particular, given, measurement
outcome plays center stage, whereas the distinction between biorthogonal versus
non-biorthogonal decompositions has dropped to an auxiliary role.
So like Schrödinger, one could argue as follows: suppose we have the wave func-
tion (25.1), considered in their general argument. Suppose we make a measurement
of quantity A on system 1, whose eigenfunctions are |ai , and eigenvalues ai . If this
measurement yields the outcome ak we infer that the corresponding wave function
for sytem 2 is ψ k . This leads us to predict with 100% probability, before making
any measurement on system 2 that the physical quantity P k has the value 1. By
554 J. Uffink

the “criterion of reality”, this entails that in this case P k is an element of reality.
However, this conclusion is conditional on the outcome obtained on system 1, i.e.
the value ak obtained in a measurement of A on system 1. And, even if system 2 is
non-interacting with, or even spatially separated from system 1, we cannot conclude
that the entire variety (for k = 1, 2, . . . of wave functions |ψ k  or associated
quantities P k refer to the same physical reality of system 2, precisely because they
do not depend only on what measurement is performed on system 1 but also on the
outcome obtained.
However, contra Schrödinger, one could continue as follows: suppose the
procedure mentioned in the previous paragraph is in place.
We perform the measurement A on system 1 and obtain one of its possible outcomes ak . For
each of those specific outcomes one of the quantities P k represents an element of reality for
system 2. So, if we do not care what this outcome might be, we can infer that at least one of
the quantities P 1 , P 2 , . . . represents a element of physical reality.

In other words, by quantifying over all possible outcomes of the A measurement,

we obtain a logical disjunction of

[P 1 = 1] ∨ [P 2 = 1] · · · ∨ [P n = 1] (25.7)

and one could regard this description as representing an element of reality for the
second system. Now compare this to another measurement on system 1, of the
quantity A . This would lead similarly to a logical disjunction of

[Q1 = 1] ∨ [Q2 = 1] · · · ∨ [Qn = 1] (25.8)

where Qi = |φ i φ i |.
But in general, of course, these two disjunctions will be mutually exclusive, and
indeed, there will be no description within quantum mechanics for a state in which
both disjunctions are correct descriptions. So, assuming with EPR that what is real
for system 2 cannot depend on the choice of measurement on system 1 we do reach
their conclusion that QM is incomplete, even for non-biorthogonal wave functions.
What emerges from this discussion is thus that Schrödinger’s insistence that the
reliance on single measurements in EPR’s general argument is not sufficient to
establish their conclusion is well-founded. However, his insistence that biorthogonal
decompositions are needed for this purpose seems too strong. What I have tried to
argue is that, while they of course simplify the argument quite considerably, they are
not essential to their argument.

25.4 Einstein’s Reply to Schrödinger

Einstein’s June 19 1935 reply to Schrödinger’s letter has become rather well-
known, since it has been quoted and discussed by several commentators (Fine 1986,
25 Schrödinger’s Reaction to the EPR Paper 555

2017; Howard 1985, 1990; Bacciagaluppi and Crull 2020). In this letter, Einstein
expresses his dissatisfaction with the formulation of the EPR paper, blaming
Podolsky for “smothering it in erudition”. It is also the first time he explicitly puts
forward a ‘Separation Principle’ (Trennungsprinzip), which might be paraphrased
by saying that for two spatially separated systems, each has its own physical state,
independently of what is done to the other system.
Einstein first presents a simple classical illustration: a ball is always found in
either of two boxes, and we describe this situation by saying that the probability of
finding it in the first or second box is 12 . He asks whether this description could be
considered as a complete specification, and contemplates two answers:
No: A complete specification is: the ball is in the first box (or not). This is how the
characterization of the state should look in the case of a complete description.
Yes: Before I open the lid of a box, the ball is not at all in one of the two boxes. Its being
in one determinate box only comes about because I open the lid. Only that brings
about the statistical character of the world of experience, [. . . ]
We have analogous alternatives when we want to interpret the relation between
quantum mechanics and reality. In the ball-system, the second “spiritist” or Schrödinger
interpretation is, so to say, silly, and only the first “Born” interpretation would be taken
seriously.8 But the talmudic philosopher whistles at “reality” as a sock puppet of naivety
and declares both views as only different in their ways of expression. [. . . ]
In quantum theory, one describes the real state of a system by means of a normalized
function ψ of its coordinates (in configuration space).[. . . ] One would now like to say the
following: ψ is correlated one-to-one with the real state of the real system. [. . . ] If this
works, I speak of a complete description of reality by the theory. But if such an interpretation
fails I call the theoretical description ‘incomplete’. [. . . ]
Let us describe the joint system consisting of subsystems [1] and [2] by a ψ-function
[|
]. This description refers to a moment in time when the interaction has practically
ceased. This ψ of the joint system can be built up from the normalized eigen-ψ |ai ,
|bj  that belong to the eigenvalues of the “observables” (respectively sets of commuting
observables) [A] and [B] respectively. We can write:

|
 = cij |ai |bj  (25.9)
ij

If we now make an [A]-measurement on 1, this expression reduces to


|ψ = cij |bj  (25.10)
j

This is the ψ-function of the subsystem [2] in the case I have made an [A]-measurement.
Now instead of the eigenfunctions of the observables [A] and [B], I can also decompose
in the eigenfunctions [of] [A ], and [B], where [A ] is another set of commuting observables:

|
 = c ij |ai |bj  (25.11)
ij

8 Here, the reference to the “Schrödinger” and “Born” interpretation refer back to the year 1926,
when Schrödinger originally interpreted the wave function as corresponding to the physical state
of the system, whereas Born proposed a statistical interpretation.
556 J. Uffink

So that after measurement of [A ] one obtains


|ψ   = c ij |bj  (25.12)
j

Ultimately, it is only essential that |ψ and |ψ   are at all different. I hold that this variety is
incompatible with the hypothesis that the ψ-description is correlated in a one-to-one fashion
with physical reality.

This reply is striking, in several ways. First of all, as has been noted by previous
authors, because the argument is significantly different from the EPR argument:
Einstein does not mention the Reality Criterium, and employs a very different notion
of completeness. The Separability Principle, although appearing here for the first
time, is arguably implicitly present in the EPR paper (Fine 1986, 2017; Howard
1985, 1990). A second reason this reply is striking is that it seems to ignore the
remarks that Schrödinger put forward in his previous letter, and with regard to
Schrödinger’s contention that biorthogonal decompositions are essential, Einstein’s
only remark is “Das ist mir wurst” (I could not care less).
Let me try to spell this out in more detail. The conclusion of Einstein’s argument
is that depending on which measurement we perform on system 1, one should assign
different wave functions to system 2, even though the real physical state of system
2 should —according to the Separation Principle— be the same, regardless of what
we choose to measure on system 1. This conclusion just reiterates the general
conclusion of the EPR paper, which was the focus of Schrödinger’s worries in the
first place. But where, for EPR, this conclusion was only an initial step —and did not
by itself imply the incompleteness of the quantum description of reality—, Einstein
argues that it is all that is ultimately essential for incompleteness.
This contrast between Einstein’s June 19 argument and the EPR argument is
due, of course, to their different construals of the notion of completeness. EPR
provide only a necessary condition for when a theoretical description is complete.
On the other hand, Einstein, in the quote above, describes a necessary and sufficient
condition in terms of a one-to-one correspondence between real states and their
quantum description by means of a wave function.
Lehner (2014) has proposed an apt terminology for dissecting notions of
completeness employed by EPR and in Einstein’s letter to Schrödinger. Failures of
a one-to-one correspondence between “real states” and a “theoretical description”
may come in at least two ways:

Overcompleteness Several theoretical descriptions are compatible with the same

real state.
Incompleteness Several real states are compatible with the same theoretical
description.

Einstein in his letter to Schrödinger is content with showing that QM is over-

complete: depending on what measurement one performs on the first subsystem,
a different wave function is attributed to the second system, even though (given
Separability) the real state of the second system should be one and the same.
25 Schrödinger’s Reaction to the EPR Paper 557

Indeed, one main reason why EPR need to continue their argument to establish
their special conclusion is just that they aim to establish incompleteness rather than
overcompleteness. (Another reason is of course their reliance on the Criterion of
Reality— I will come back to that below.)
The surprising point, here, is that, cases of overcompleteness are ubiquitous
within physics. Indeed, even in quantum theory, it is well-known that one can
multiply a wave function ψ with an arbitrary phase factor eiα (α ∈ R ) to obtain
another wave function ψ  = eiα ψ, where it is generally assumed that ψ and ψ 
describe the same physical situation. More general examples of overcompleteness
can be found in gauge theories. Indeed one might wonder whether overcompleteness
is a worrisome issue at all in theoretical physics. Now of course, in the examples just
alluded to, the gauge freedom in the theoretical description of a physical state can
usually be expresses in terms of equivalence relations and be expelled by ‘dividing
out’ the corresponding equivalence classes in the theoretical descriptions. No such
equivalence relations exist in the case discussed by Einstein in the June 19 letter,
and so the case of overcompleteness presented there in his is not so simple as in the
examples of gauge freedom (cf. Howard 1985).
But apart from the distinctions between Einstein (on June 19) and EPR, let me
return to Schrödinger. As we have seen, Schrödinger had misgivings precisely about
the EPR general conclusion, and tried to point out to Einstein that the wave function
assigned to system 2 depends not only on what measurement is performed on system
1, but also on its outcome.
Einstein’s letter, however, does not contain any acknowledgement of this issue.
For example, he argues that the wave function (25.10) “is the ψ function of the
subsystem [2] in the case I have made an [A]-measurement”. But it is not, or at the
very least, Einstein’s notation is ambiguous here. Inspection of the right-hand side
of (25.10) reveals that this wave function depends on i, which labels the outcome
obtained in the measurement of A. So a proper formulation, instead of (25.10)
should have read something like


|ψ i  = cij |bj  (25.13)
j

This is the ψ-function of subsystem 2 in case I have made an A-measurement and obtained
the outcome ai .

But this notation makes it transparent that there will be different wave functions for
system 2 for different outcomes ak and ar , and, as Schrödinger argued, this need not
be paradoxical. A similar remark holds for the wave function (25.12).
Now, one might argue for a more charitable reading of Einstein’s response of
June 19 to Schrödinger. One such reading9 is that Einstein may have wanted to
point out that that if we perform measurement A on system 1 the wave function of
system 2 would be one of the form

9I am grateful to Guido Bacciagaluppi for suggesting this reading.

558 J. Uffink


|ψ i  = cij |bj  (25.14)
j

or, more precisely, that it would be one of the wave functions in the set

{|ψ 1 , . . . |ψ n } (25.15)

On the other hand, if we perform a measurement of A on system 1, the second

system will be left in one of the wave functions

{|ψ 1  . . . |ψ n } (25.16)

and in general, these two sets will be disjoint. So, if there were a wave function to
describe the real physical state of system 2, independently of what measurement we
perform on system 1, it would have to be a member of two disjoint sets. In other
words, there would not even be any wave function to assign to system 2. And so, we
would recover Einstein’s conclusion, i.e. a failure of a one-to-one correspondence
between between physical states and wave functions, not in terms of Lehner’s
incompleteness (many-to-one) nor overcompleteness (one-to-many) but rather as
a one-to-none correspondence. Be that as it may, it would of course still be failure
of one-to-one correspondence.
However, this still leaves another vulnerability in Einstein’s argument to be
discussed. This vulnerability is that the argument will only convince those who
share Einstein’s premise that the wave function is to be interpreted as a direct
description of the real state of a system. In the June 19 letter, he ascribes this view
to Schrödinger, but Schrödinger sets him right in a response of August 19:
I have long passed the stage when I thought that one could see the ψ-function as a direct
description of reality.

Indeed, there are numerous passages in Schrödinger’s post-1927 work that indicate
he came to acknowledge that the wave function contained epistemic factors,
pertaining to our knowledge about the system, as well as objective factors pertaining
to its physical state. But perhaps worse than Schrödinger’s rejection of this premise
is that adherents of the Copenhagen interpretation, and in particular the ‘talmudic
philosopher’ (Bohr), i.e. the main target of the argument, will also reject it.
In this regard, the EPR paper itself fares much better. The reason is that it does
not attempt to establish a link between physical reality and a wave function, but
only between elements of physical reality and measurement outcomes of physical
quantities, at least when they are fully predictable. How does this change the
argument? In the case of their special conclusion, one could argue as follows:
assume the Separability Principle. But depending on what quantity we measure on
system 1 (either P or Q), we can infer with certainty the values of the corresponding
measurements on system 2. They should therefore represent elements of physical
reality, indeed simultaneously so, because what is real on system 2 should, by
Separability, not depend on what measurement we perform on system 1. But there
25 Schrödinger’s Reaction to the EPR Paper 559

is no quantum description of a state with definite real values for both position an
momentum for system 2. Therefore the quantum description of physical reality is
incomplete (in the sense of a one-to-none correspondence).
The reason why the EPR special conclusion is more robust against objections,
compared to Einstein’s June 19 argument, is precisely because it avoids the premise
that wave functions directly describe physical reality. But this comes at a price:
the EPR special conclusion employs the maximally entangled EPR state with it
its multiple biorthogonal decompositions, i.e. the assumption that Einstein said he
could not care less about. However, as I have argued in the previous section, this
conclusion can be generalized to a case without such biorthogonal ecompositions.

25.5 What Did Schrödinger Make of the EPR Argument?

Schrödinger was in 1935 no stranger to the EPR argument—having worked out

the basics of the example in 1931—but unlike EPR, or Einstein in his 1935
correspondence to Schrödinger, he never expressed his conclusions to the argument
in terms of (in)completeness.

25.5.1 Schrödinger’s Interpretation of the Wave Function

Einstein, in his June 19 letter ascribes the view that the wave function presents a
complete description of physical reality to Schrödinger. This view is indeed one
that Schrödinger used in his original works on wave mechanics in the first half of
1926 (Joas and Lehner 2009). However, he abandoned this view in several letters
written in November 1926, precisely because of the phenomenon we would now
call entanglement.
Indeed, in his (1935b), he writes that the wave function is only a “representative”
of the state of the system, and that under the proviso that he is describing the current
view (Lehrmeinung) of quantum mechanics. More specifically, he characterizes a
wave function as a mere catalogue of expectation values, and compares it to a travel
guide (Baedeker). All this underlines that Schrödinger did not commit to a view that
the wave function was in one-to-one correspondence with the real state of a system.

25.5.2 Schrödinger’s Take on the EPR Problem

Einstein persistently and consistently saw the EPR problem as an objection to the
claim that the quantum description of physical systems by means of a wave function
was complete, and also persistently argued that this problem would disappear if
only one would take the view of the wave function as a statistical ensemble (thus
560 J. Uffink

providing an incomplete description of a physical system). He made this argument

in another letter to Schrödinger of August 8:
My solution to the paradox presented in our [EPR] paper is this: the ψ-function does
not describe the state of a single system, but rather gives a (statistical) description of an
ensemble of systems.

Schrödinger, by contrast, did not harbour any such hopes. For Schrödinger,
the argument pointed to a deeper problem, that would not be resolved by merely
adopting an ensemble interpretation. Thus, in response to Einstein on August 19, he
points out that Einstein’s solution would not work. He provides an argument, using
2
the example of the operator Pa 2 + a 2 Q2 which had been presented to Schrödinger
for a similar purpose by Pauli in a letter of July 9 (von Meyenn et al. 1985). His
(1935b) contains a similar argument. Einstein’s response (on September 4) to this
argument is
If I claimed that I have understood your last letter, that would be a fat lie (von Meyenn 2011,
p. 569).

In reply, (October 4) Schrödinger apologizes and uses the example of the operator
for angular momentum, also suggested to him by Pauli on July 9, and also discussed
in (1935b). Both arguments are actually versions of Von Neumann’s argument
against a hidden variables interpretation.
But if Schrödinger did not believe in Einstein’s favorite way to resolve the
paradox, what alternative did he have? It seems to me that Schrödinger did not
see a clear way out. For example in his letter to Sommerfeld (December 1931)
Schrödinger concluded
One can hardly help oneself in any other way than by the following ‘emergency decree’: in
quantum mechanics it is forbidden to make statements about what is ‘real’, about the object,
such statements deal only with the relation object–subject, and apparently in a much more
restrictive way than in any other description of nature. (von Meyenn 2011, Vol. I, p. 490)

This suggests that Schrödinger saw the only rescue in a non-realist interpretation of
quantum mechanics. In the Übungen notebook, his conclusion was a rejection of the
Copenhagen interpretation;
Therefore: This interpretational scheme does not work. One has to look for another one.

But in his June 7 letter to Einstein, he put blame on the non-relativist nature of QM:
The lesson (Der Vers) that I have drawn for myself until now about this whole business
is this. We do not possess a quantum mechanics compatible with relativity theory, that
is, amongst others, a theory that takes account of the finite speed of propagation of all
interactions. [. . . ] When we want to separate two systems, their interaction will, long before
it vanishes, fail to approximate the absolute Coulomb or similar force. And that is where
our understanding ends. The process of separation cannot at all be grasped in the orthodox
scheme.

Similarly, his final remarks in (1935b) point out the non-relativistic nature of
QM, and the obstacles towards a fully relativistic theory, and conjecture that non-
relativistic QM might merely be a convenient calculational artifice. For Schrödinger,
25 Schrödinger’s Reaction to the EPR Paper 561

the conclusions to be drawn form the EPR argument were therefore not so clear-
cut. On some occasions he would be satisfied by just claiming that it showed
an inadequacy of the Copenhagen interpretation of quantum mechanics, on other
occasions he hoped that a future relativistic version of quantum theory would make
the problem disappear, but without having concrete proposals about how this would
work.

25.6 The Origins of the Cat Paradox

Schrödinger’s long essay, Die gegenwärtige Situation in der Quantenmechanik

appeared in die Naturwissenschaften in three installments in subsequent issues in
November and December of 1935. His famous cat paradox made its entry in the
second installment, published on November 29, 1935. There is a widespread view
in the literature (Fine 1986, 2017; Moore 1992; Gribbin 2012; Gilder 2008; Kaiser
2016) that Schrödinger only conceived of his famous cat paradox in response to
his correspondence with Einstein over the summer of 1935. Indeed, Einstein wrote
to Schrödinger on August 8 about an example of an explosive device, containing
an amount of gunpowder in a chemically unstable state. Its wave function would
then develop in the course of time to contain parts that represent an exploded
and un unexploded state. This example is indeed, in all intent and purposes, very
similar to Schrödinger’s cat. Given the lapse between August 8 and November
29, it might seem, on first sight, very well possible that Schrödinger adapted this
example of Einstein into his own paper. For example, in his authoritative biography
of Schrödinger, (Moore 1992, p. 305) comments that “Einstein’s gunpowder would
soon reappear in the form of Schrödinger’s cat.” Similarly, (Gribbin 2012, p. 179–
180) writes: “The ideas encapsulated in the famous ‘thought experiment’ involving
Schrödinger’s cat actually came in no small measure from Einstein, in the extended
correspondence between the two triggered by the EPR paper.”
However, if we look a bit closer into the genesis of Schrödinger (1935b),
this view becomes more problematic. Arnold Berliner, founder and editor of die
Naturwissenschaften invited several authors in June 1935 to comment on the EPR
paper, amongst them Pauli, Heisenberg, and Schrödinger. And although Pauli and
Heisenberg did consider this invitation seriously, and Heisenberg indeed wrote a
manuscript for this purpose, (Bacciagaluppi and Crull 2020), only Schrödinger
submitted a contribution. Berliner wrote to Schrödinger on July 1, to express his
joy that Schrödinger had consented to contribute to a discussion about EPR in die
Naturwissenschaften. He addresed the question about whether “the name” could be
mentioned in Schrödinger’s contribution.10 Berliner emphasizes that Schrödinger
should feel free to mention the Innominato in a footnote or otherwise would be

10 AsVon Meyenn (p. 548) notes, this issue concerns the fact that under the Nazi regime, the
mention of Einstein’s name in German scientific literature had been forbidden.
562 J. Uffink

entirely up to his discretion. In a letter of July 13 to Einstein, Schrödinger mentions

that he had begun writing such a paper:
The great difficulty for me to reach even only a mutual understanding with the orthodox
people, has led me to make an attempt to analyse the present interpretational situation ab
ovo in a longish essay. Whether and what I will publish of this, I do not know.

The letter continues by describing some of the themes that did indeed end up in the
pubished (Schrödinger 1935b).
On July 25, Schrödinger wrote to Berliner that his manuscript was finished in
handwriting, and that a citation of the work from the Physical Review would only
appear after three quarters of his manuscript, at the beginning of the final quarter. He
mentions that it would take him some more time to produce a type-written version.
He writes to Berliner again on August 11, saying he did not reply earlier:
because the manuscript, which you so kindly agreed to publish, was doubly burning under
my fingers. I have now put it all together, and will send it off tomorrow together with this
letter.

Apparently, his paper, eventually published as (1935b) in November and December,

was finished in typewriting on August 11, and sent off the next day.
Now, it is obvious that Einstein’s August 8 letter with the gun-powder example,
written from Old Lyme (Connetticut), could not have crossed the ocean to reach
Schrödinger, in Oxford, before he sent off his manuscript on August 12. Indeed,
in his reply (August 19) to Einstein’s August 8 letter, Schrödinger states that in a
long essay he had just sent off a week ago he had made use of an example, very
similar to Einstein’s gun powder example and describes his cat example to Einstein.
Thus, the only sensible conclusion at this point can be that Einstein and Schrödinger
conceived of their strikingly similar examples independently, around the same time,
cf. Ryckman (2017).
Still, the most elaborate study and commentary of the Schrödinger-Einstein
correspondence in 1935 to date is (Fine 1986) which includes a chapter called
“Schrödinger’s Cat and Einstein’s: the genesis of a paradox”, in which he discussed
the question how Schrödinger might have conceived of his famous cat paradox.
This work is presumably the origin for the wide-spread conception that that the
Cat Paradox borrowed and built upon Einstein’s exploding gunpowder example of
August 8.
Indeed, Fine claims that
The cat was put into (Schrödinger 1935b) in response to ideas and distinctions generated
during the correspondence [with Einstein] over EPR (p. 73).

But Fine’s argument differs in detail from Moore and Gribbin. He suggests that
section 4 and also section 5 (containing the Cat) of (1935b) were not part of the
paper sent to Berliner on August 12, but later additions:
The conclusion seems inescapable that section 4 was put together only after the August 19
letter [by Schrödinger], and was not part of the original manuscript [Schrödinger] sent to
Berliner. (p. 81)
25 Schrödinger’s Reaction to the EPR Paper 563

But if section 4 was only written after August 19, then most likely section 5 is a later
addition as well. (p. 82)

There are several problems with this hypothesis. I will deal with its first claim
(i.e. that section 4 of Schrödinger’s (1935b) could only have been written after
Schrödinger’s August 19 letter) below. Let me first mention problems with the
conjecture that section 5, containing the Cat Paradox, was composed later as
well. The first problem is that it would imply that Schrödinger is literally lying
to Einstein when he wrote on August 19, in response to Einstein’s exploding
gunpowder example, that he had already presented a quite similar example in the
paper he just sent off a week earlier, and described the cat example. This is all the
more improbable given that Schrödinger seems to be generally generous in issues
of priority, and did not even emphasize in his June 7 letter that the EPR paper
“about that which we discussed so much in Berlin” is in fact indebted to his own
formulation of the EPR problem in 1931.
Of course, Fine’s hypothesis could be tested, e.g. if a draft version of the Natur-
wissenschaften paper surfaces. I could not find any in the Schrödinger Nachlass,
(but who knows what might emerge from archives of Die Naturwissenschaften, if
they survived the war). However, one can rely on what is known from Schrödinger’s
correspondence in this period with Berliner.
When Schrödinger wrote to Berliner on July 25 that he had finished a handwritten
draft of the paper, and would still need time to type it. He reported a word
count of 10230 words, and estimated that this would fill 13.3 to 14 pages of die
Naturwissenschaften not counting the paragraph headings. (von Meyenn 2011, p.
555). He suggested to that this “extraordinary length” probably required it to be cut
into two or three installments, (but advising against the latter option).
On July 29, Berliner replied that he was looking forward to the submission, and
that he would happily allot all the space Schrödinger needed, and promised that it
would appear in two subsequent issues of the journal.
On August 11, Schrödinger wrote to Berliner that his typescript was finished and
to be mailed the next day. On August 14, Berliner responded that he had suddenly
been removed from his position as editor of die Naturwissenschaften, but implored
Schrödinger that his paper should still be published in that journal, and that he had
proposed Walther Matthée as his successor.11
On August 19, Schrodinger wrote Einstein, as mentioned before, and told him
also that his first impulse was to withdraw the paper, but didn’t because Berliner
himself had insisted on its publication.
On August 24, he received a letter from Mathée, telling Schrödinger he was
the new editor of Die Naturwissenschaften, thanking him for the manuscript,
and mentions including the proofs of the paper. Mathée informed him that the

11 Berliner
was removed from the journal because he was Jewish (von Meyenn 2011, Vol. 2,
pp. 546–7).
564 J. Uffink

manuscript was to be split into three parts, to appear in consecutive issues, and
asked Schrödinger to indicate in these proofs where the cuts were to be made.12
Schrödinger must have replied to Mathée soon afterwards, and there is a draft of
this letter in the selection by Von Meyenn, but erroneously appended to his letter
to Berliner from July 25. (von Meyenn 2011, Vol. 2, p. 557). Here, Schrödinger
mentions that he had just sent a telegram with the message “Sorry, tripartion
impossible” and maintains that the manuscript should only be split into two parts,
or otherwise returned to him. He mentions that this division could well take place
after section 9, which is “very exactly half of the manuscript”. He ends, however, on
a more reconciliatory note, saying that he trusts Mathée’s judgment, who should not
feel bound by an earlier promise, and that Schrödinger would feel the same way if
Berliner had told him that the promise of a bipartition was not possible. There is no
mention at all of a revised version with later additions to the paper at this stage, and
clearly there would have been very little time to do so, if this revision is supposed
to have occurred after August 19.
In fact, the paper was published in three installments in Schrödinger (1935b). In
total, they comprise about 12270 words (not counting footnotes, references, figure
captions, tables of contents and section titles)13 and span about 15.5 pages. This
implies that Schrödinger did expand his manuscript by about 2000 extra words,
or about 2 pages, after July 25. This could actually fit Fine’s hypothesis that
its sections 4 and 5 (which together comprise just less than 2 pages, and about
1600 words) could be later additions — but not necessarily additions from after
August 19.
The second problem for Fine’s hypothesis concerns Schrödinger’s estimate on
July 25 that the first mention of Einstein’s name would occur after three quarters of
the paper, as well as his estimate, presumably a few days after August 24, that the
end of section 9 would “ very exactly” divide the manuscript into two equal halves.
Both these estimates are actually quite accurate also for the published version. This
suggests that, more likely, he made additions while typing his draft of July 25, more
or less equally spaced throughout the manuscript. In any case, if Schrödinger added
two whole new sections before section 9 in Fine’s hypothetical revision one must
surmise him to have also made additions of the same size to the second half of his
manuscript as well in order to keep the end of section 9 at exactly half of the paper.
But that clearly did not happen, because that would take the size of the manuscript
beyond what was actually published.
Given this evidence, it seems quite unlikely that Schrödinger only thought of
the cat paradox after he read Einstein’s exploding device example. The more
simple and straightforward explanation, as stated earlier, is that these two authors,
independently, thought of very similar examples at roughly the same time.
Next, let me discuss Fine’s argument for the “inescapable” claim that section 4
of Schrödinger (1935b) is a later addition to the manuscript Schrödinger sent to

12 This letter is not included in the selection by Von Meyenn.

13 I am indebted to Guido Bacciagaluppi for this careful word count.
25 Schrödinger’s Reaction to the EPR Paper 565

Berliner on August 12. This claim is based on the fact that Einstein writes in a
letter to Schrödinger on August 9, that he believes a mere ensemble interpretation
of the wave function (which, of course would concede that the description by a
wave function is incomplete (in Einstein’s sense as well as the EPR sense of the
term) would suffice to solve the paradox. Schrödinger responds on August 19 that
this argument will not work. He proceeds to provide a counter-argument, using the
Hamiltonian of the harmonic oscillator as an example. Section 4 of Schrödinger
(1935b) contains a very similar argument.
Fine assumes that Schrödinger only could have developed this argument in
response to Einstein’s letter, and thus after August 19. However, it seems to me
that this assumption overlooks Schrödinger’s wider correspondence in this period.
The very idea that an interpretation of the wave function as incomplete, i.e., as
merely describing a statistical inhomogeneous ensemble is irreconcilable with the
quantum predictions was mentioned to Schrödinger by Pauli in a letter of July 9
1935, specifically pointing out that the discrete spectrum that quantum mechanics
requires for operators like the Hamiltonian P 2 + ω2 Q2 for a harmonic oscillator
or the angular momentum (in a 3-dimensional case) Px Qy − Py Qx would be
incompatible with this view. Schrödinger apparently took Pauli’s remarks to heart.
Both the harmonic oscillator Hamiltonian and the angular momentum examples
are discussed in Schrödinger (1935b), with his own twist, adding an arbitrary
multiplicative constant to the last term of the harmonic oscillator Hamiltonian.
Moreover, he presents this argument in a version adapted to the EPR case in a
letter to von Laue (July 25 1935). Again, I believe there seems little ground for
the hypothesis that Schrödinger could only have composed this argument after he
responded to Einstein on August 19, and even less for the claim that, therefore,
section 4 of (1935b) must have been a later addition to the manuscript he submitted
on August 12 of that year.
Indeed, it seems to me that Fine’s analysis of Schrödinger’s Naturwissenschaften
paper, vis-a-vis his correspondence with Einstein in 1935, paints him as a passive
receiver of Einstein’s ideas, and as composing his own (Schrödinger 1935b) paper
merely in response to those ideas by Einstein. I believe that we should redress this
picture. Schrödinger knew about the essential aspects of the EPR argument 3 12 years
before it was published, and might well have influenced Einstein in his transition
from the photon box argument to the clearer and simpler version that is EPR. There
was no need for Schrödinger to learn about EPR from Einstein. Instead, as I have
argued, he had a lesson to teach Einstein, although he failed to get across.

Acknowledgements I am indebted to The Max Planck Institute for History of Science in Berlin
for enabling this research, as well as the Vossius Institute in Amsterdam and the Descartes Centre
in Utrecht for providing the opportunity to finish it. I also thank Christoph Lehner for many
valuable discussions, and deciphering Schrödinger’s handwriting, and Guido Bacciagaluppi for
many helpful and insightful comments. I am also grateful to Alexander Zartl for assistance in
browsing through the Schrödinger Nachlass in the Zentralbibliothek für Physik at the University
of Vienna.
566 J. Uffink

References

Bacciagaluppi, G., & Crull E. (2020). The Einstein paradox: The debate on nonlocality and
incompleteness in 1935. Cambridge: Cambridge University Press.
Einstein, A., Podolsky, B., & Rosen, N. (1935). Can quantum mechanical description of reality be
considered complete? Physical Review, 47, 777–780.
Fine, A. (1986). The Shaky game: Einstein, realism and the quantum theory. Chicago: University
of Chicago Press.
Fine, A. (2017). The Einstein-Podolsky-Rosen argument in quantum theory. In E. N. Zalta (Ed) The
Stanford encyclopedia of philosophy (Winter 2017 Edition). https://plato.stanford.edu/archives/
win2017/entries/qt-epr/.
Gilder, L. (2008). The age of entanglement. New York: Knopf.
Gribbin, J. (2012). Erwin Schrödinger and the quantum revolution. London: Bantam Press.
Howard, D. (1985). Einstein on locality and separability. Studies in History and Philosophy of
Science, 16, 171–201.
Howard, D. (1990). Nicht sein kann was nicht sein darf, or the Prehistory of EPR, 1909–1935:
Einstein’s early worries about the Quantum Mechanics of Composite Systems. In A. I. Miller
(Ed.) Sixty-two years of uncertainty: Historical, philosophical and physical inquiries in the
foundations of quantum mechanics (NATO ASI series, Vol. 226, pp. 61–111). New York:
Plenum Press.
Joas, C. & Lehner, C. (2009). The classical roots of wave mechanics: Schrödinger’s transfor-
mations of the optical-mechanical analogy. Studies in the History and Philosophy of Modern
Physics, 40, 338–351.
Kaiser, D. (2016). How Einstein and Schrödinger conspired to kill a cat. Nautilus, Issue 041.
Lehner, C. (2014). Einstein’s realism and his critique of quantum mechanics. In M. Janssen & C.
Lehner (Eds) The Cambridge companion to Einstein (pp. 306–353). Cambridge: Cambridge
University Press.
Moore, W. (1992). Schrödinger, life and thought. Cambridge: Cambridge University Press.
Ryckman, T. (2017). Einstein. Milton Park: Routledge.
Schrödinger, E. (1935a). Discussion of probability relations between separated systems. Mathe-
matical Proceedings of the Cambridge Philosophical Society, 31, 555–563.
Schrödinger, E. (1935b). Die gegenwärtige Situation in der Quantenmechanik. Die Naturwis-
senschaften, 23, 807–812, 823–828, 844–848.
von Meyenn, K. (2011). Eine entdeckung von ganz außerordentliche Tragweite; Schrödinger’s
Briefwechsel zur Wellenmechanik und zum Katzenparadoxon (Vols. 1 and 2). Heidelberg:
Springer.
von Meyenn, K., Hermann, A., & Weiskopf, V. F. (1985). Wolfgang Pauli: Wissenschaftlicher
Briefwechsel mit Bohr, Einstein, Heisenberg u.a. Vol II: 1930–1939. Heidelberg: Springer.
Chapter 26
Derivations of the Born Rule

Lev Vaidman

Abstract The Born rule, a cornerstone of quantum theory usually taken as a

postulate, continues to attract numerous attempts for its derivation. A critical review
of these derivations, from early attempts to very recent results, is presented. It is
argued that the Born rule cannot be derived from the other postulates of quantum
theory without some additional assumptions.

Keywords Quantum probability · Born rule · Many-worlds interpretation ·

Relativistic causality

26.1 Introduction

My attempt to derive the Born rule appeared in the first memorial book for Itamar
Pitowsky (Vaidman 2012). I can only guess Itamar’s view on my derivation from
reading his papers (Pitowsky 1989, 2003, 2006; Hemmo and Pitowsky 2007).
It seems that we agree which quantum features are important, although our
conclusions might be different. In this paper I provide an overview of various
derivations of the Born rule. In numerous papers on the subject I find in depth
analyses of particular approaches and here I try to consider a wider context that
should clarify the status of the derivation of the Born rule in quantum theory. I
hope that it will trigger more general analyses which finally will lead to a consensus
needed for putting foundations of quantum theory on solid ground.
The Born rule was born at the birth of quantum mechanics. It plays a crucial
role in explaining experimental results which could not be explained by classical
physics. The Born rule is known as a statement about the probability of an outcome
of a quantum measurement. This is an operational meaning which corresponds to

L. Vaidman ()
Raymond and Beverly Sackler School of Physics and Astronomy, Tel-Aviv University,
Tel-Aviv, Israel
e-mail: vaidman@post.tau.ac.il

© Springer Nature Switzerland AG 2020 567

M. Hemmo, O. Shenker (eds.), Quantum, Probability, Logic,
Jerusalem Studies in Philosophy and History of Science,
https://doi.org/10.1007/978-3-030-34316-3_26
568 L. Vaidman

numerous very different statements about ontology in various interpretations of

quantum theory. Von Neumann’s description of quantum measurement includes, at
some stage, a collapse of the quantum state corresponding to a particular result of
the measurement and the Born rule provides the probability of getting this result.
In this framework there is no definition of when exactly it happens (where is
the quantum-classical boundary?), so the Copenhagen interpretation and its recent
development in the form of QBism (Caves et al. 2002) do not specify the ontology,
leaving the definition of the principle on the operational level. In the framework
of the Bohmian interpretation (Bohm 1952), which is a deterministic theory given
the initial Bohmian positions of all particles, the Born rule is a postulate about the
probability distribution which governs these random initial Bohmian positions. It is
a postulate about the genuinely random stochastic collapse process in the framework
of physical collapse theories (Ghirardi et al. 1986). In Aharonov’s solution to the
measurement problem (Aharonov et al. 2014), it is a postulate about the particular
form of the backward evolving wavefunciton. In the framework of the many-worlds
interpretation (MWI) (Everett III 1957), it is a postulate about experiences of an
observer in a particular world (Vaidman 2002). So, in all interpretations, the Born
rule is postulated, but the question of the possibility of its derivation is considered
to be of interest, and it is still open (Landsman 2009).
A rarely emphasized important fact about the Born rule is that it might be
even more relevant for explaining physics phenomena in which the probability is
not explicitly manifested. Quantum statistical mechanics which leads to quantum
thermodynamics heavily uses the Born rule for explaining everyday observed
phenomena. There is nothing random when we observe a blue sky or reddish sun
at sunset. The explanation includes scattered photons of various colors absorbed by
cones in the eye with their color depended efficiency of the absorption. The ratio of
the number of events of photon absorption in different cones corresponds to different
experiences of the color of the sky and the sun, and we have to use the Born rule
to explain our visual experience (Li et al. 2014). In this explanation we consider
the cone photoreceptor in an eye as a single-photon detector and light from the sun
scattered by molecules of air as a collection of photons. The quantum nature of
light coming from modern artificial light sources is even more obvious. An observer
looks on a short flash of a fluorescent soft white bulb and announces the color she
has seen. The spectrum of the light from this bulb consists of red, green and blue
photons, but nobody would say that she saw red light or that she saw green light from
the fluorescent bulb. The Born Rule is needed to calculate the ratio of the signals
from the cones. The large number of events of photon detection by cones explains
why nobody would say they saw a color different from white, since the Born rule
provides an almost vanishing probability for such event. The Born rule also tells us
that there is an astronomically small probability that we will see a red sky and a blue
sun, but it is not different from other quantum tiny tails which we neglect when we
explain the classical world we observe through underlying quantum reality.
26 Derivations of the Born Rule 569

26.2 Frequentist Approach

One of the early approaches relied on consideration of infinite series of repeated

measurements. In the frequentist approach to probability, we consider the ratio of
particular outcomes to the total number of measurements. The probability acquires
its meaning only in the infinite limit. The important milestones were the works
of Hartle (1968) and Farhi et al. (1989). Then the program was extended by
replacing infinite tensor products of Hilbert spaces by continuous fields of C$ -
algebras (Van Wesep 2006; Landsman 2008). The core feature of these arguments
involves taking a limit of an infinite number of quantum systems. Aharonov and his
collaborators (Aharonov and Reznik 2002; Aharonov et al. 2017) presented, in my
view, the simplest and the most elegant argument based on this type of infinite limit.
Consider a large number N of identical systems all prepared in the same state

|Ψ  = α i |ai , (26.1)
i

which is a superposition of nondegenerate

 eigenstates of a variable A. Consider the
“average” variable Ā ≡ N1 N n=1 An . Applying the universal formula (Aharonov
and Vaidman 1990)

A|Ψ  = Ψ |A|Φ |Ψ  + ΔA|Ψ⊥ , (26.2)

where Ψ |Ψ⊥  = 0, we obtain

%
N %
N
ΔA  %
N
Ā |Ψ n = Ψ |A|Ψ  |Ψ n + |Ψ n |Ψ⊥ k . (26.3)
N
n=1 n=1 k=1 n=k

The amplitude of the first term in the right hand side of the equation is of order
1 while the amplitude of the second term (the sum) is proportional to √1 , so in
N
the limit
N as N tends to infinity, the second term can be neglected and the product
state n=1 |Ψ n can be considered an eigenstate of the variable Ā with eigenvalue
Ψ |A|Ψ .
Now consider the measurement of Ā followed by measurements of A of each
of the individual systems. Ni is the number of outcomes A = ai . The probability
of outcome ai is defined as the limit pi ≡ limN →∞ NNi . To derive the Born rule
we consider the shift of the pointer of the measuring device measuring Ā, in two
ways. First, since in the limit, the state is an eigenstate with eigenvalue Ψ |A|Ψ ,
the pointer is shifted by this value. Second, consider the evolution backward in time
given that we have the results of  individual measurements of variable A of each
system. Then the shift has to be N a i Ni
i=1 N . In the limit we obtain Ψ |A|Ψ  =
 N
i |α i | ai = i=1 ai pi . This equation can be generally true only if pi = |α i | for
2 2

all eigenvalues ai . This proves the Born rule.

570 L. Vaidman

The legitimacy of going to the limit N → ∞ in the earlier proofs was questioned
in Squires (1990), Buniy et al. (2006) and Aharonov’s approach was analyzed in
Finkelstein (2003). I am also skeptical about the possibility of arguments relying on
the existence of infinities to shed light on Nature. Surely, the infinitesimal analysis
is very helpful, but infinities lead to numerous very peculiar sophisticated features
which we do not observe. I see no need for infinities to explain our experience.
Very large numbers can mimic everything and are infinitely simpler than infinity.
The human eye cannot distinguish between 24 pictures per second and continuous
motion, but infinite information is required to describe the latter. There is no need
for infinities to explain all what we see around.
Another reason for my skepticism about possibility to understand Nature by
neglecting vanishing terms in the infinite limit is the following example in which
these terms are crucial for providing common sense explanation. In the modification
of the interaction-free measurement (Elitzur and Vaidman 1993) based on the Zeno
effect (Kwiat et al. 1995) we get information about the presence of an object without
being near it. The probability of success can be made as close to 1 as we wish
by changing the parameters. Together with this, there is an increasing number of
moments of time at which the particle can be absorbed by the object with decreasing
probability of each absorption. In the limit, the sum of the probabilities of absorption
at all different moments goes to zero, but without these cases the success of the
interaction-free measurement seems to contradict common sense. These are the
cases in which there is an interaction. Taking the limit in proving the Born rule
is analogous to neglecting these cases.
The main reason why I think that this approach cannot be the solution is that I
do not see what is the additional assumption from which we derive the Born rule.
Consider a counter example. Instead of the Born rule, the Nature has “Equal rule”.
Every outcome ai of a measurement of A has the same probability given that it
is possible, i.e., α i = 0 . Of course, this model contradicts experimental results,
but it does not contradict the unitary evolution part of the formalism of quantum
mechanics. I do not see how making the number of experiments infinite can rule
out Equal rule. Note that an additional assumption ruling out this model is hinted in
Aharonov and Reznik (2002) “the results of physical experiments are stable against
small perturbations”. A very small change of the amplitude can make a finite change
in the probability in the proposed model. (This type of continuity assumption is
present in some other approaches too.)

26.3 The Born Rule and the Measuring Procedure

The Born rule is intimately connected to the measurement problem of quantum

mechanics. Today there is no consensus about its solution. The Schrödinger equation
cannot explain definite outcomes of quantum measurements. So, if it does not
explain the existence of a (unique) outcome, how can it explain its probability?
It is the collapse process (which is not explained by the Schrödinger equation) that
26 Derivations of the Born Rule 571

provides the unique outcome, so it seems hopeless to look for an explanation of the
Born rule based on the Schrödinger equation.
What might be possible are consistency arguments. If we accept the Hibert
space structure of quantum mechanics and we accept that there is probability for
an outcome of a quantum measurement, what might this probability measure be?
Itamar Pitowsky suggested to take it as the basis and showed how Gleason’s theorem
(Gleason 1957) (which has its own assumptions) leads to the Born rule (Pitowsky
1998). He was aware of “two conceptual assumptions, or perhaps dogmas. The
first is J. S. Bell’s dictum that the concept of measurement should not be taken
as fundamental, but should rather be defined in terms of more basic processes.
The second assumption is that the quantum state is a real physical entity, and that
denying its reality turns quantum theory into a mere instrument for predictions”
(Bub and Pitowsky 2010). In what followed, he recognized the problem as I do:
“This last assumption runs very quickly into the measurement problem. Hence,
one is forced either to adopt an essentially non-relativistic alternative to quantum
mechanics (e.g. Bohm without collapse, GRW with it); or to adopt the baroque
many worlds interpretation which has no collapse and assumes that all measurement
outcomes are realized.” The difference between us is that he viewed “the baroque
many worlds interpretation” as unacceptable (Hemmo and Pitowsky 2007), while I
learned to live with it (Vaidman 2002).
Maybe more importantly, we disagree about the first dogma. I am not ready to
accept “measurement” as a primitive. Physics has to explain all our experiences,
from observing results of quantum measurements to observing the color of the sun
and the sky at sunset. I do believe in the ontology of the wave function (Vaidman
2016, 2019) and I am looking for a direct correspondence between the wave function
and my experience considering quantum observables only as tools for helping to find
this correspondence. I avoid attaching ontological meaning to the values of these
observables. It does not mean that I cannot discuss the Born rule. The measurement
situation is a well defined procedure and our experiences of this procedure (results
of measurements) have to be explained.
The basic requirement of the measurement procedure is that if the initial state
is an eigenstate of the measured variable, it should provide the corresponding
eigenvalue with certainty. Any procedure fulfilling this property is a legitimate
measuring procedure. The Born rule states that the probability it provides should
be correct for any legitimate procedure, and this is a part of what has to be proved,
but let us assume that the fact that all legitimate procedures provide the same
probabilities is given. I will construct then a particular measurement procedure
(which fulfills the property of probability 1 for eigenstates) which will allow me
to explain the probability formula of the Born rule.
Consider a measurement of variable A performed on a system prepared in the
state (26.1). The measurement procedure has to include coupling to the measuring
device and the amplification part in which the result is written in numerous quantum
systems providing a robust record. Until this happens, there is no point in discussing
the probability, since outcomes were not created yet. So the measurement process
is:
572 L. Vaidman

%  % %
|Ψ  |rMD
m → α i |ai  |MD
m |rMD
m , m r|m
MD MD
= 0 ∀m
m i m∈Si m∈S
/ i
(26.4)
where “ready” states |rMD m of the numerous parts m of the measuring device,
m ∈ Si , are changed to macroscopically different (and thus orthogonal) states
|MD
m , in correspondence with the eigenvalue ai . For all possible outcomes ai , the
set Si of subsystems of the measuring device which change their states to orthogonal
states has to be large enough. This part of the process takes place according to all
interpretations. In collapse interpretations, at some stage, the state collapses to one
term in the sum.
This schematic description is not too far from reality. Such a situation appears
in a Stern-Gerlach experiment in which the atom, the spin component of which is
measured, leaves a mark on a screen by exciting numerous atoms of the screen.
But I want to consider a modified measurement procedure. Instead of a screen in
which the hitting atom excites many atoms, we put arrays of single-atom detectors
in the places corresponding to particular outcomes. The arrays cover the areas of
the quantum uncertainty of the hitting atom. The arrays are different, as they have
a different number Ni of single-atom detectors which we arrange according to
equation Ni = |α i |2 N . (We assume that we know the initial state of the system.)
In the first stage of the modified measuring procedure an entangled state of the
atom and sensors of the single-atom detectors is created:
%  αi  %
|Ψ  |rsen
n → √ |ai  |sen
ki |rsen
n , (26.5)
n i
N i k n=k
i i

where |rsen
n represents an unexcited state of the sensor with label n running
over sensors of all arrays of single-photon detectors. N is the total number of
detectors. For each eigenvalue ai there is one array of Ni detectors with sensors
in a superposition of entangled states in which one of the sensors ki has an excited
state |sen
ki . At this stage the measurement has not yet taken place. The number of
sensors with changed quantum state might be large, but no “branches” with many
systems in excited states has been created. We need also the amplification process
which consists of excitation of a large number of subsystems of individual detectors.
In the modified measurement, instead of a multiple recording of an event specified
by the detection of ai , we record activation of every sensor ki by excitation of a
large (not necessarily the same) number of quantum subsystems m belonging to
the set Sik . Including in our description these subsystems, the description of the
measurement process is:
% % 1   % % %
|Ψ  |rsen
n |rMD
m → √ |ai  |sen
ki |rsen
n |MD
m |rMD
m .
n m N i k n=k m∈S m∈S
/
i i ik ik
(26.6)
Here we also redefined the states |sen
to absorb the phase of α i to see explicitly
ki
that all terms in the superposition have the same amplitude. Every term in the
26 Derivations of the Born Rule 573

superposition has macroscopic number of subsystems of detectors with states which

are orthogonal to the states appearing in other terms. This makes all the terms
separate. We have N different options. They consist of sets according to all different
possible eigenvalues when the set corresponding to eigenvalue ai has Ni elements.
Assuming that all options are equiprobable, we obtain the Born rule. The probability
of a reading corresponding to eigenvalue ai is pi = NNi = |α i |2 . And this procedure
is a good measurement according to our basic requirement: If the initial state is an
eigenstate, we will know it with certainty.
In Fig. 26.1, we demonstrate such a √ situation for√ a modified Stern-Gerlach
experiment with the initial state |Ψ  = 0.4|↑ + 0.6|↓. There are N = 5
single-photon detectors. The description of the measurement process (represented
for a general case by [26.6]) is now:
√ √ %
5 %
0.4|↑ + 0.6|↓ |rsen
n |rMD
m →
n=1 m
& '
1 % % % % %
√ |↑|sen
1 +|↑|sen
2 +|↓|sen
3 +|↓|sen
4 +|↓|sen
5 ,
5 1 2 3 4 5

(26.7)
where
% % % %
≡ |rsen
n |MD
m |rMD
m .
i n=i m∈Si m∈S
/ i

We obtain a superposition of five equal-amplitude states, each corresponding to one

detector clicks and others are not. It is natural to accept equal probabilities for clicks
of all these detectors and since there are two detectors corresponding to the outcome
‘up’ and three detectors corresponding to the outcome ‘down’ we obtain the Born
rule probabilities for our example.
An immediate question is: how can I claim to derive pi = |α i |2 when in my
procedure I put in by hand NNi = |α i |2 ? The answer is that making another choice
would not lead to a superposition of orthogonal terms with equal amplitudes, so
with another choice the derivation does not go through.
This derivation makes a strong assumption that in the experiment, the firing of
each sensor has the same probability. It is arranged that all these events correspond
to terms in the superposition with the same amplitude, so the assumption is that
equal amplitudes correspond to equal probabilities. It is this fact that is considered
to be the main part in the derivation of the Born rule. I doubt that the formalism
of quantum mechanics by itself is enough to provide a proof for this statement, see
also Barrett (2017). In the next section I will try to identify the assumptions added
in various proofs of the Born rule.
Without the proof of the connection between amplitudes and probabilities, the
analysis of the experiment I presented above is more of an explanation of the Born
rule than its derivation. We also use an assumption that all valid measurement
574 L. Vaidman

√ √
Fig. 26.1 Modified Stern-Gerlach experiment specially tailored for the state 0.4|↑ + 0.6|↓.
There are two detectors in the location corresponding to the result ‘up’ and three detectors in the
location corresponding to the result ‘down’

experimental procedures provide the same probabilities for outcomes. The modified
procedure has a very natural combinatorial counting meaning of probability. It can
be applied to the collapse interpretations when we count possible outcomes and in
the MWI where we count worlds. The objection that the number of worlds is not a
well defined concept (Wallace 2010) is answered when we put weights (measures
of existence) on the worlds (Vaidman 1998; Greaves 2004; Groisman et al. 2013).

26.4 Symmetry Arguments

In various derivations of the Born rule, the statement that equal amplitudes lead to
equal probabilities relies on symmetry arguments. The starting point is the simplest
(sometimes named pivotal) case

1
|Ψ  = √ (|a1  + |a2 ). (26.8)
2

The pioneer in attempting to solve this problem was Deutsch (1999) whose work
was followed by extensive development by Wallace (2007), a derivation by Zurek
26 Derivations of the Born Rule 575

(2005), and some other attempts such as Sebens and Carroll (2016) and also my
contribution with McQueen (Vaidman 2012; McQueen and Vaidman 2018). The
key element of these derivations is the symmetry under permutation between |a1 
and |a2 . It is a very controversial topic, with numerous accusations of circularity
for some of the proofs (Hemmo and Pitowsky 2007; Barnum et al. 2000; Saunders
2004; Gill 2005; Schlosshauer and Fine 2005; Lewis 2010; Rae 2009; Dawid and
Thébault 2014).
In all these approaches there is a tacit assumption (which I also used above)
that the probability of an outcome of a measurement of A does not depend on
the procedure we use to perform this measurement. Another assumption is that
probability depends only on the quantum state. In the Deutsch-Wallace approach
some manipulations, swapping, and erasures are performed to eliminate the dif-
ference between |a1  and |a2 , leading to probability half due to symmetry. If the
eigenstates do not have internal structure except for being orthogonal states, then
symmetry can be established, but it seems to me that these manipulations do not
provide the proof for important realistic cases in which the states are different in
many respects. It seems that what we need is a proof that all properties, except
for amplitudes, are irrelevant. I am not optimistic about the existence of such a
proof without adding some assumptions. Indeed, what might rule out the “Equal
rule”, a naive rule according to which probabilities for all outcomes corresponding
to nonzero amplitudes are equal, introduced above?
The assumption of continuity of probabilities as functions of time rules Equal
rule out, but this is an additional assumption. The Deutsch-Wallace proof is in
the framework of the MWI, i.e. that the physical theory is just unitary evolution
which is, of course, continuous, but it is about amplitudes as functions of time. The
experience, including the probability of self-location of an observer, supervenes
on the quantum state specified by these amplitudes, but the continuity of this
supervenience rule is not granted.
Zurek made a new twist in the derivation of the Born rule (Zurek 2005). His key
idea is to consider entangled systems and rely on “envariance” symmetry. A unitary
evolution of a system which can be undone by the unitary evolution of the system it
is entangled with. For the pivotal case, the state is

1
|Ψ  = √ (|a1 |1 + |a2 |2), (26.9)
2

where |1, |2 are orthogonal states of the environment. The unitary swap
|a1  ↔ |a2  followed by the unitary swap of the entangled system |1 ↔ |2 brings
us back to the original state which, by assumption, corresponds to the original
probabilities. Other Zurek’s assumptions are that a manipulation of the second
system does not change the probability of the measurement on the system, while
the swap of the states of the systems swaps the probabilities for the two outcomes.
This proves that the probabilities for the outcomes in the pivotal example must be
equal.
576 L. Vaidman

In my view, the weak point is the claim that swapping the states of the system
swaps the probabilities of the outcomes. This property follows from the quantum
formalism when the initial state is an eigenstate, but in our case, when the
mechanism for the choice of the outcome is unknown, we also do not know how
it is affected by unitary operations. Note, that it is not true for the Stern-Gerlach
experiment in the framework of Bohmian mechanics.
Zurek, see also Wallace (2010), Baker (2007), and Boge (2019), emphasises the
importance of decoherence: entanglement of environment with eigenstates of the
system. Indeed, decoherence is almost always present in quantum measurements
and its presence might speed up the moment we can declare that the measurement
has been completed, but, as far as I understand, decoherence is neither necessary,
nor sufficient for completing a quantum measurement. It is not necessary, because it
is not practically possible to perform an interference experiment with a macroscopic
detector in macroscopically different states even if it is isolated from the environ-
ment. It is not sufficient, because decoherence does not ensure collapse and does not
ensure splitting of a world. For a proper measurement, the measuring device must
be macroscopic. It is true that an interaction of the system with an environment,
instead of a macroscopic measuring device, might lead to a state similar to (26.4)
with macroscopic number of microscopic systems of environment “recording” the
eigenvalue of the observable. It, however, does not ensure that the measurement
happens. It is not clear that macroscopic number of excited microsystem causes a
collapse, see analysis of a such situation in the framework of the physical collapse
model (Ghirardi et al. 1986) in Albert and Vaidman (1989) and Aicardi et al. (1991).
In the framework of the many-worlds interpretation we need splitting of worlds. The
moment of splitting does not have a rigorous definition, but a standard definition
(Vaidman 2002) is that macroscopic objects must have macroscopically different
states. Decoherence might well happen due to a change of states of air molecules
which do not represent any macroscopic object.
What I view as the most problematic “symmetry argument proof” of probability
half for the pivotal example is the analysis of Sebens and Carroll (2016), see also
Kent (2015). Sebens and Caroll considered the measurement in the framework of the
MWI and apply the uncertainty of self-location in a particular world as a meaning
of probability (Vaidman 1998). However, in my understanding of the example
they consider, this uncertainty does not exist (McQueen and Vaidman 2018). In
their scenario, a measurement of A on a system in state (26.8) is performed on a
remote planet. Sebens and Caroll consider a question: What is the probability of
an observer who is here, i.e., far away from the planet, to be in a world with a
particular outcome? This question is illegitimate, because he is certainly present in
both worlds, there is no uncertainty here. This conclusion is unavoidable in the MWI
as I understand it (Vaidman 2002), which is a fully deterministic theory without any
room for uncertainty. However, uncertainty in the MWI is considered in Saunders
(2004) and Saunders and Wallace (2008), so if this program succeeds (see however
Lewis 2007), then the Sebens-Caroll proof might make sense. Another way to make
26 Derivations of the Born Rule 577

sense of the Sebens-Caroll proof was proposed by Tappenden (2017) based on

his unitary interpretation of mind, but I have difficulty accepting this metaphysical
picture.
A scenario in which an observer is moved to different locations according to
an outcome of a quantum measurement without getting information about this
outcome (Vaidman 1998), allows us to consider the probability based on observer’s
ignorance about self-location and without uncertainty in the theory. This by itself,
however, does not prove the probability half for the pivotal case. The proof
(McQueen and Vaidman 2018), which is applicable to all interpretations, has two
basic assumptions. First, it is assumed that space in Nature has symmetry, so we
can construct the pivotal case with symmetry between the states |a1  and |a2 . We
do not rely on permutation of states, we rely on the symmetry of physical space
and construct a symmetric state with identical wave packets in remote locations
1 and 2. The second assumption is that everything fulfills the postulate of the
theory of special relativity according to which we cannot send signals faster than
light. Changing probability by a remote action is sending signals. This proves that
changing the shape or even splitting a remote state will not change the probability
of finding a1 provided its amplitude was not changed.

26.5 Other Approaches

Itamar Pitowsky’s analysis of the Born rule on the basis of Gleason’s theorem
(Pitowsky 1998) was taken further to the case of generalized measurements (Caves
et al. 2004). Galley and Masanes (2017) continued research which singles out
the Born rule from other alternatives. Note that they also used symmetry (“bit
symmetry”) to single out the Born rule. Together with Muller, they extended their
analysis (Masanes et al. 2019) and claimed to prove everything just from some
“natural” properties of measurements which are primitive elements in their theory.
So, people walked very far on the road paved by Itamars’s pioneering works. I
have to admit that I am not sympathetic to this direction. The authors of Masanes
et al. (2019) conclude “Finally, having cleared up unnecessary postulates in the
formulation of QM, we find ourselves closer to its core message.” For me it seems
that they go away from physics. Quantum mechanics was born to explain physical
phenomena that classical physics could not. It was not a probability theory. It was
not a theory of measurements, and I hope it will not end as such. “Measurements”
should not be primitives, they are physical processes as any other, and physics
should explain all of them.
Similarly, I cannot make much sense of claims that the Born rule appears even
in classical systems presented in the Hilbert space formalism (Brumer and Gong
2006; Deumens 2019). Note that in the quantum domain, the Born rule appears
even outside the framework of Hilbert spaces in the work of Saunders (2004),
who strongly relies on operational assumptions such as a continuity assumption:
“sufficiently small variations in the state-preparation device, and hence of the initial
578 L. Vaidman

state, should yield small variations in expectation value.” This assumption is much
more physical than postulates of general probabilistic theories.
The dynamical derivation in the framework of the Bohmian interpretation cham-
pioned by Valentini (Valentini and Westman 2005; Towler et al. 2011) who argued
that under some (not too strong) requirements of complexity, the Born distribution
arises similarly to thermal probabilities in ordinary statistical mechanics. See
extensive discussion in Callender (2007) and recent analysis in Norsen (2018) which
brings also similar ideas from Dürr et al. (1992). The fact that for some initial
conditions of some systems relaxation to Born statistics does not happen is a serious
weakness of this approach. What I find more to the point as a proof of the Born rule
is that the Born statistical distribution remains invariant under time evolution in all
situations. And that, under some very natural assumptions, is the only distribution
with this strong property (Goldstein and Struyve 2007).
Wallace (2010) and Saunders (2010) advocate analyzing the issue of probability
in the framework of the consistent histories approach. It provides formal expressions
which fit the probability calculus axioms. However, I have difficulty seeing what
these expressions might mean. I failed to see any ontological meaning for the main
concept of the approach “the probability of a history”, and it also has no operational
meaning apart from the conditional probability of an actually performed experiment
(Aharonov et al. 1964), while the approach is supposed to be general enough to
describe evolution of systems which were not measured at the intermediate time.

26.6 Summary of My View

I feel that there is a lot of confusion in the discussions of the subject and it is
important to make the picture much more clear. Even if definite answers might not
be available now, the question: What are the open problems? can be clarified. First,
it is important to specify the framework: collapse theory, hidden variables approach
or noncollapse theory. Although in many cases the “derivation of the Born rule”
uses similar structure and arguments in all frameworks, the conceptual task is very
different. I believe that in all frameworks there is no way to prove the Born rule
from other axioms of standard quantum mechanics. The correctly posed question is:
What are the additional assumptions needed to derive the Born rule?
Standard quantum mechanics tells us that the evolution is unitary, until it leads
to a superposition of quantum states corresponding to macroscopically different
classical pictures. There is no precise definition of “macroscopically different
classical pictures” and this is a very important part of the measurement problem,
but discussions of the Born rule assume that this ambiguity is somehow solved, or
proven irrelevant. The discussions analyze a quantitative property of the nonunitary
process which happens when we reach this stage assuming that the fact that it
happens is given. I see no possibility to derive the quantitative law of this nonunitary
process from laws of unitary evolution. It is usually assumed that the process
depends solely on the quantum state, i.e. that the probability of an outcome of
26 Derivations of the Born Rule 579

a measurement of an observable does not depend on some hidden variables and

does not depend on the way the observable is measured. The process also should
not alter the unitary evolution when a superposition of states corresponding to
macroscopically different classical pictures was not created. This, however, is not
enough to rule out proposals different from the Born rule, e.g., Equal rule described
above. We have to add something to derive the Born rule. We are not supposed
to rely on experimental results, they do single out the Born rule, but this is not a
“derivation”. Instead, if we take some features of observed results as the basis, it is
considered as a derivation. I am not sure that it is really better, unless these features
are considered not as properties of Nature, but as a basic reason for Nature to be as
it is. Then the Born rule derivations become a part of the program to get quantum
mechanics from simple axioms (Popescu and Rohrlich 1994; Hardy 2001; Chiribella
et al. 2011). In these derivations, quantum mechanics is usually considered as a
general probability theory and the main task is to derive the Born rule.
In McQueen and Vaidman (2018) the program is more modest. Unitary quantum
mechanics is assumed and two physical postulates are added. First, that there are
symmetries in space and second that there is no superluminal signalling. The first
principle allows us to construct a pivotal example described by (26.8) in which there
is symmetry between states |a1  and |a2 . The second principle allows us to change
one of the eigenstates in the pivotal state without changing the probability to find
the other eigenvalue. This is the beginning of the procedure, first shown by Deutsch
in (1999), who pioneered these types of derivations.
The situation in the framework of the MWI is conceptually different. The
physical essence of the MWI is: unitary evolution of a quantum state of the universe
is all that there is. There is no additional process of collapse behaviour which should
be postulated. So it seems that here there is no room for additional assumptions and
that the Born rule must be derived just from the unitary evolution.
However, the MWI has a problem with probability even before we discuss the
quantitative formula of the Born rule. The standard approach to the probability of
an event requires that there to be a matter of fact about whether this event and not
the other takes place, but in the MWI all events take place. On the other hand, we do
have experience of one particular outcome when we perform a measurement. My
resolution of this problem (Vaidman 1998) is that indeed, there is no way to ask
what is the probability of what will happen, because all outcomes will be actual.
The “probability” rule is still needed to explain statistics of observed measurements
in the past. There are worlds with all possible statistics, but we happen to observe
Born rule statistics. The “probability” explaining these statistics is the probability
of self-location in a particular world. In Vaidman (1998) I constructed a scenario
with quantum measurements in which the observer is split (and together with him,
his world) according to the outcome of the measurement without being aware of the
result of the measurement. This provides the ignorance probability of the observer
about the world specified by the outcome of the measurement he is a part of.
Tappenden (2010) argues that merely considering such a construction allows us to
discuss the Born rule. These are supporting arguments of the solution: there is no
probabilistic process in Nature: with certainty all possible outcomes of a quantum
580 L. Vaidman

measurement will be realized, but an observer, living by definition in one of the

worlds, can consider the question of probability of being located in a particular
world.
All that is in Nature, according to MWI, is a unitary evolving quantum state of
the universe and observers correspond to parts of this wave function (Vaidman 2016,
2019). So, there is a hope that the experience of observers, after constructing the
theory of observers (chemistry, biology, psychology, decision theory, etc.) can be,
in principle, explained solely from the evolution of the quantum state. Apparently,
experiences of an observer can be learned from his behavior which is described
by the evolution of the wave function. Then, it seems that the Born rule should be
derivable from the laws of quantum mechanics. However, I believe that this is not
true.
Consider Alice and Bob at separate locations and they have a particle in a
state (26.8) where |a1  corresponds to a particle being at Alice’s site and |a2 
corresponds to a particle being at Bob’s site. Now assume that instead of the Born
rule, which states that the probability of self-location in a world is proportional to the
square of the amplitude, Nature has the Equal rule which yields the same probability
of self location in all the worlds, i.e. probability N1 , where N is the number of worlds.
Equal rule allows superluminal signaling. Alice and Bob agree that at a particular
time t Alice measures the presence of the particle at her site, i.e. she measures the
projection on state |a1 . To send bit 0, Bob does nothing. Alice’s measurement splits
the world into two worlds: the one in which she finds the particle and the other, in
which she does not. Then she has equal probability to find herself in each of the
worlds, so she has probability half to find herself in the world in which she finds the
particle. For sending bit 1, just before time t, Bob performs a unitary operation on
the part of the wave at his site splitting it to a hundred orthogonal states


100
|a2  → |bk , (26.10)
k=1

and immediately measures operator B which tells him the eigenvalue bk . This
operation splits the initial single world with the particle in a superposition into
hundred and one worlds: hundred worlds with one of Bob’s detectors finding the
particle, and one world in which the particle was not found by Bob’s detectors.
Prior to her measurement, Alice is present in all these worlds. Her measurement
1
tests if she is in one particular world, so she has only probability of 101 to find the
particle at her site.
Bob’s unitary operation and measurement change the probability of Alice’s
outcome. With measurements on a single particle, the communication is not very
reliable, but using an ensemble will lead to only very rare cases of an error. The
Equal rule will ensure that Alice and Bob meeting in the future will (most probably)
verify correctness of the bit Bob has sent.
We know that unitary evolution does not allow superluminal communication.
(When we consider a relativistic generalisation of the Schrödinger equation.) Can,
26 Derivations of the Born Rule 581

a supertechnology, capable of observing superposition of Alice’s and Bob’s worlds,

given that the actual probability rule of self-location is the Equal rule (and not the
Born rule), use the above procedure for sending superluminal signals? No! Only
Alice and Bob, inside their worlds, have the ability of superluminal communication.
It does not contradict relativistic properties of physics describing unitary evolution
of all worlds together.
What I argue here, is that the situation in the framework of MWI is not different
from collapse theories. There is a need for an independent probability postulate.
In collapse theory it is a physical postulate telling us about the dynamics of the
ontology, dynamics of the quantum state of the universe describing the (single)
world. In MWI, the postulate belongs to the part connecting observer’s experiences
with the ontology. In the MWI, as in a collapse theory, the experiences supervene
on the ontology, the quantum state. The supervenience rule is the same when the
quantum state corresponds to a single world, but it has an additional part regarding
the probability of self-location, when the quantum state of the universe corresponds
to more than world. The postulate describes this supervenience rule.
We can justify the Born rule postulate of self-location by experimental evidence,
or by requiring the relativistic constraint of superluminal signaling also within
worlds. I find a convincing explanation in the concept of the measure of existence
of a world (Vaidman 1998; Groisman et al. 2013). While there is no reason to
postulate that the probability of self-location in every world is the same, it is
natural to postulate that the probability of self-location in worlds of equal existence
(equal square of the amplitude) is the same. Adding another natural assumption
that probability of self-location in a particular world should be equal to the sum of
the probability of self-location in all the worlds which split from the original one,
provides the Born rule.
My main conclusion is that there is no way to derive the Born rule without
additional assumptions. It is true both in the framework of collapse theories and,
more surprisingly, in the framework of the MWI. The main open question is not the
validity of various proofs, but what are the most natural assumptions we should add
for proving the Born rule.

Acknowledgements This work has been supported in part by the Israel Science Foundation Grant
No. 2064/19.

References

Aharonov, Y., & Reznik, B. (2002). How macroscopic properties dictate microscopic probabilities.
Physical Review A, 65, 052116.
Aharonov, Y., & Vaidman, L. (1990). Properties of a quantum system during the time interval
between two measurements. Physical Review A, 41, 11–20.
Aharonov, Y., Bergmann, P. G., & Lebowitz, J. L. (1964). Time symmetry in the quantum process
of measurement. Physical Review, 134, B1410–B1416.
582 L. Vaidman

Aharonov, Y., Cohen, E., Gruss, E., & Landsberger, T. (2014). Measurement and collapse within
the two-state vector formalism. Quantum Studies: Mathematics and Foundations, 1, 133–146.
Aharonov, Y., Cohen, E., & Landsberger, T. (2017). The two-time interpretation and macroscopic
time-reversibility. Entropy, 19, 111.
Aicardi, F., Borsellino, A., Ghirardi, G. C., & Grassi, R. (1991). Dynamical models for state-vector
reduction: Do they ensure that measurements have outcomes? Foundations of Physics Letters,
4, 109–128.
Albert, D. Z., & Vaidman, L. (1989). On a proposed postulate of state-reduction. Physics letters A,
139, 1–4.
Baker, D. J. (2007). Measurement outcomes and probability in Everettian quantum mechanics.
Studies in History and Philosophy of Science Part B: Studies in History and Philosophy of
Modern Physics, 38, 153–169.
Barnum, H., Caves, C. M., Finkelstein, J., Fuchs, C. A., & Schack, R. (2000). Quantum probability
from decision theory? Proceedings of the Royal Society of London. Series A: Mathematical,
Physical and Engineering Sciences, 456, 1175–1182.
Barrett, J. A. (2017). Typical worlds. Studies in History and Philosophy of Science Part B: Studies
in History and Philosophy of Modern Physics, 58, 31–40.
Boge, F. J. (2019). The best of many worlds, or, is quantum decoherence the manifestation of
a disposition? Studies in History and Philosophy of Science Part B: Studies in History and
Philosophy of Modern Physics, 66, 135–144.
Bohm, D. (1952). A suggested interpretation of the quantum theory in terms of “hidden” variables.
I. Physical Review, 85, 166.
Brumer, P., & Gong, J. (2006). Born rule in quantum and classical mechanics. Physical Review A,
73, 052109.
Bub, J., & Pitowsky, I. (2010). Two dogmas about quantum mechanics. In S. Saunders, J. Barrett,
A. Kent, & D. Wallace (Eds.), Many Worlds?: Everett, quantum theory, & reality (pp. 433–459).
Oxford: Oxford University Press.
Buniy, R. V., Hsu, S. D., & Zee, A. (2006). Discreteness and the origin of probability in quantum
mechanics. Physics Letters B, 640, 219–223.
Callender, C. (2007). The emergence and interpretation of probability in Bohmian mechanics.
Studies in History and Philosophy of Science Part B: Studies in History and Philosophy of
Modern Physics, 38, 351–370.
Caves, C.M., Fuchs, C.A., & Schack, R. (2002). Quantum probabilities as Bayesian probabilities.
Physical Review A, 65, 022305.
Caves, C. M., Fuchs, C. A., Manne, K. K., & Renes, J. M. (2004). Gleason-type derivations of the
quantum probability rule for generalized measurements. Foundations of Physics, 34, 193–209.
Chiribella, G., D’Ariano, G. M., & Perinotti, P. (2011). Informational derivation of quantum theory.
Physical Review A, 84, 012311.
Dawid, R., & Thébault, K. P. (2014). Against the empirical viability of the Deutsch–Wallace–
Everett approach to quantum mechanics. Studies in History and Philosophy of Science Part B:
Studies in History and Philosophy of Modern Physics, 47, 55–61.
Deumens, E. (2019). On classical systems and measurements in quantum mechanics. Quantum
Studies: Mathematics and Foundations, 6, 481–517.
Deutsch, D. (1999). Quantum theory of probability and decisions. Proceedings of the Royal Society
of London. Series A: Mathematical, Physical and Engineering Sciences, 455, 3129–3137.
Dürr, D., Goldstein, S., & Zanghi, N. (1992). Quantum equilibrium and the origin of absolute
uncertainty. Journal of Statistical Physics, 67, 843–907.
Elitzur, A. C., & Vaidman, L. (1993). Quantum mechanical interaction-free measurements.
Foundations of Physics, 23, 987–997.
Everett III, H. (1957). “Relative state” formulation of quantum mechanics. Reviews of Modern
Physics, 29, 454–462.
Farhi, E., Goldstone, J., & Gutmann, S. (1989). How probability arises in quantum mechanics.
Annals of Physics, 192, 368–382.
26 Derivations of the Born Rule 583

Finkelstein, J. (2003). Comment on “How macroscopic properties dictate microscopic probabili-

ties”. Physical Review A, 67, 026101.
Galley, T. D., & Masanes, L. (2017). Classification of all alternatives to the Born rule in terms of
informational properties. Quantum, 1, 15.
Ghirardi, G.C., Rimini, A., & Weber, T. (1986). Unified dynamics for microscopic and macroscopic
systems. Physical Review D, 34, 470–491.
Gill, R. (2005). On an argument of David Deutsch. In M. Schürmann & U. Franz (Eds.), Quantum
probability and infinite dimensional analysis: From foundations to applications (QP-PQ Series,
Vol. 18, pp. 277–292). Singapore: World Scientific.
Gleason, A. M. (1957). Measures on the closed subspaces of a Hilbert space. Journal of
Mathematics and Mechanics, 6, 885–893.
Goldstein, S., & Struyve, W. (2007). On the uniqueness of quantum equilibrium in Bohmian
mechanics. Journal of Statistical Physics, 128, 1197–1209.
Greaves, H. (2004). Understanding Deutsch’s probability in a deterministic multiverse. Studies
in History and Philosophy of Science Part B: Studies in History and Philosophy of Modern
Physics, 35, 423–456.
Groisman, B., Hallakoun, N., & Vaidman, L. (2013). The measure of existence of a quantum world
and the Sleeping Beauty Problem. Analysis, 73, 695–706.
Hardy, L. (2001). Quantum theory from five reasonable axioms. arXiv preprint quant-ph/0101012.
Hartle, J. B. (1968). Quantum mechanics of individual systems. American Journal of Physics, 36,
704–712.
Hemmo, M., & Pitowsky, I. (2007). Quantum probability and many worlds. Studies in History
and Philosophy of Science Part B: Studies in History and Philosophy of Modern Physics, 38,
333–350.
Kent, A. (2015). Does it make sense to speak of self-locating uncertainty in the universal wave
function? Remarks on Sebens and Carroll. Foundations of Physics, 45, 211–217.
Kwiat, P., Weinfurter, H., Herzog, T., Zeilinger, A., & Kasevich, M. A. (1995). Interaction-free
measurement. Physical Review Letters, 74, 4763–4766.
Landsman, N. P. (2008). Macroscopic observables and the Born rule, I: Long run frequencies.
Reviews in Mathematical Physics, 20, 1173–1190.
Landsman, N. P. (2009). Born rule and its interpretation. In D. Greenberger, K. Hentschel, &
F. Weinert (Eds.), Compendium of Quantum Physics: Concepts, Experiments, History and
Philosophy (pp. 64–70). Heidelberg: Springer.
Lewis, P. J. (2007). Uncertainty and probability for branching selves. Studies in History and
Philosophy of Science Part B: Studies in History and Philosophy of Modern Physics, 38, 1–
14.
Lewis, P. J. (2010). Probability in Everettian quantum mechanics. Manuscrito: Revista Interna-
cional de Filosofía, 33, 285–306.
Li, P., Field, G., Greschner, M., Ahn, D., Gunning, D., Mathieson, K., Sher, A., Litke, A., &
Chichilnisky, E. (2014). Retinal representation of the elementary visual signal. Neuron, 81,
130–139.
Masanes, L., Galley, T. D., & Müller, M. P. (2019). The measurement postulates of quantum
mechanics are operationally redundant. Nature Communications, 10, 1361.
McQueen, K. J., & Vaidman, L. (2018). In defence of the self-location uncertainty account of
probability in the many-worlds interpretation. Studies in History and Philosophy of Science
Part B: Studies in History and Philosophy of Modern Physics, 66, 14–23.
Norsen, T. (2018). On the explanation of Born-rule statistics in the de Broglie-Bohm pilot-wave
theory. Entropy, 20, 422.
Pitowsky, I. (1989). Quantum probability-quantum logic. Berlin: Springer.
Pitowsky, I. (1998). Infinite and finite Gleason’s theorems and the logic of indeterminacy. Journal
of Mathematical Physics, 39, 218–228.
Pitowsky, I. (2003). Betting on the outcomes of measurements: A Bayesian theory of quantum
probability. Studies in History and Philosophy of Science Part B: Studies in History and
Philosophy of Modern Physics, 34, 395–414.
584 L. Vaidman

Pitowsky, I. (2006). Quantum mechanics as a theory of probability. In Physical theory and its
interpretation (pp. 213–240). Berlin/Heidelberg: Springer.
Popescu, S., & Rohrlich, D. (1994). Quantum nonlocality as an axiom. Foundations of Physics,
24, 379–385.
Rae, A. I. (2009). Everett and the Born rule. Studies in History and Philosophy of Science Part B:
Studies in History and Philosophy of Modern Physics, 40, 243–250.
Saunders, S. (2004). Derivation of the Born rule from operational assumptions. Proceedings of
the Royal Society of London. Series A: Mathematical, Physical and Engineering Sciences, 460,
1771–1788.
Saunders, S. (2010). Chance in the Everett interpretation. In S. Saunders, J. Barrett, A. Kent, &
D. Wallace (Eds.), Many worlds?: Everett, quantum theory, & reality (pp. 181–205). Oxford:
Oxford University Press.
Saunders, S., & Wallace, D. (2008). Branching and uncertainty. The British Journal for the
Philosophy of Science, 59, 293–305.
Schlosshauer, M., & Fine, A. (2005). On Zurek’s derivation of the Born rule. Foundations of
Physics, 35, 197–213.
Sebens, C. T., & Carroll, S. M. (2016). Self-locating uncertainty and the origin of probability in
Everettian quantum mechanics. The British Journal for the Philosophy of Science, 69, 25–74.
Squires, E. J. (1990). On an alleged “proof” of the quantum probability law. Physics Letters A,
145, 67–68.
Tappenden, P. (2010). Evidence and uncertainty in Everett’s multiverse. British Journal for the
Philosophy of Science, 62, 99–123.
Tappenden, P. (2017). Objective probability and the mind-body relation. Studies in History and
Philosophy of Science Part B: Studies in History and Philosophy of Modern Physics, 57, 8–16.
Towler, M., Russell, N., & Valentini, A. (2011). Time scales for dynamical relaxation to the Born
rule. Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences,
468, 990–1013.
Vaidman, L. (1998). On schizophrenic experiences of the neutron or why we should believe in
the many-worlds interpretation of quantum theory. International Studies in the Philosophy of
Science, 12, 245–261.
Vaidman, L. (2002). Many-Worlds interpretation of Quantum mechanics. In E. N. Zalta (Ed.), The
Stanford encyclopedia of philosophy. Metaphysics Research Lab, Stanford University. https://
plato.stanford.edu/cgi-bin/encyclopedia/archinfo.cgi?entry=qm-manyworlds
Vaidman, L. (2012). Probability in the many-worlds interpretation of quantum mechanics. In
Y. Ben-Menahem & M. Hemmo (Eds.), Probability in physics (pp. 299–311). Berlin: Springer.
Vaidman, L. (2016). All is ψ. Journal of Physics: Conference Series, 701, 012020.
Vaidman, L. (2019). Ontology of the wave function and the many-worlds interpretation. In
O. Lombardi, S. Fortin, C. López, & F. Holik (Eds.), Quantum worlds: Perspectives on the
ontology of quantum mechanics. (pp. 93–106). Cambridge: Cambridge University Press.
Valentini, A., & Westman, H. (2005). Dynamical origin of quantum probabilities. Proceedings of
the Royal Society A: Mathematical, Physical and Engineering Sciences, 461, 253–272.
Van Wesep, R. A. (2006). Many worlds and the appearance of probability in quantum mechanics.
Annals of Physics, 321, 2438–2452.
Wallace, D. (2007). Quantum probability from subjective likelihood: Improving on Deutsch’s proof
of the probability rule. Studies In History and Philosophy of Science Part B: Studies In History
and Philosophy of Modern Physics, 38, 311–332.
Wallace, D. (2010). How to prove the Born rule. In S. Saunders, J. Barrett, A. Kent, D. Wallace
(Eds.) Many worlds?: Everett, quantum theory, & reality (pp. 227–263). Oxford University
Press.
Zurek, W. H. (2005). Probabilities from entanglement, Born’s rule pk = |ψ k |2 from envariance.
Physical Review A, 71, 052105.
Chapter 27
Dynamical States and the
Conventionality of (Non-) Classicality

Alexander Wilce

Abstract Itamar Pitowsky, along with Jeff Bub and others, long championed the
view that quantum mechanics (QM) is best understood as a non-classical probability
theory. This idea has several attractions, not least that it allows us even to pose the
question of why quantum probability theory has the rather special mathematical
form that it has—a problem that has been very fruitful, and to which we now
have several rather compelling answers, at least for finite-dimensional QM. Here,
however, I want to offer some modest caveats. One is that QM is best seen, not as a
new probability theory, but as something narrower, namely, a class of probabilistic
models, selected from within a much more general framework. It is this general
framework that, if anything, deserves to be regarded as a “non-classical” probability
theory.
Secondly, however, I will argue that, for individual probabilistic models, and even
for probabilistic theories (roughly, classes of such models), the distinction between
“classical” and a “non-classical” is largely a conventional one, bound up with the
question of what one means by the state of a system. In fact, for systems with a high
degree of symmetry, including quantum mechanics, it is possible to interpret general
probabilistic models as having a perfectly classical probabilistic structure, but an
additional dynamical structure: states, rather than corresponding simply to prob-
ability measures, are represented as certain probability measure-valued functions
on the system’s symmetry group, and thus, as fundamentally dynamical objects.
Conversely, a classical probability space equipped with reasonably well-behaved
family of such “dynamical states” can be interpreted as a generalized probabilistic
model in a canonical way. It is noteworthy that this “dynamical” representation is
not a conventional hidden-variables representation, and the question of what one
means by “non-locality” in this setting is not entirely straightforward.

A. Wilce ()
Department of Mathematical Sciences, Susquehanna University, Selinsgrove, PA, USA
e-mail: wilce@susqu.edu

© Springer Nature Switzerland AG 2020 585

M. Hemmo, O. Shenker (eds.), Quantum, Probability, Logic,
Jerusalem Studies in Philosophy and History of Science,
https://doi.org/10.1007/978-3-030-34316-3_27
586 A. Wilce

Keywords Probabilistic theories · Classical representations · Dynamical states ·

Simplices · Test spaces

27.1 Introduction

Itamar Pitowsky long championed the idea (also associated with Jeff Bub, among
others) that quantum theory is best viewed as a non-classical probability theory
(Pitowsky 2005). This position is strongly motivated by the fact that the mathe-
matical apparatus of quantum mechanics can largely be reduced to the statement
that the binary observables—the physically measurable, {0, 1}-valued quantitites—
associated with a quantum system are in one-to-one correspondence with the
projections in a von Neumann algebra, in such a way that commuting families
of projections can be jointly measured and mutually orthogonal projections are
mutually exclusive.1 In other words, mathematically, quantum theory is what one
gets when one replaces the Boolean algebras of classical probability theory with
projection lattices. It is very natural to conclude from this that quantum theory
simply is a “non-classical” or “non-commutative” probability theory.
This point of view has many attractions, not least that it more or less dissolves
what Pitowsky called the “big” measurement problem (Bub and Pitowsky 2012;
Pitowsky 2005). It also raises the question of why quantum mechanics has the
particular non-classical structure that it does—that is, why projection lattices of von
Neumann algebras, rather than something more general? The past decade has seen
a good deal of progress on this question, at least as it pertains to finite-dimensional
quantum systems (Hardy 2001; Masanes and Müller 2011; D’Ariano et al. 2011;
Mueller et al. 2014; Wilce 2012).
Here, I want to offer some modest caveats. The first is that how we understand
Pitowski’s thesis depends on how we understand probability theory itself. On
the view that I favor, quantum mechanics (henceforth, QM) is best seen, not as
replacement for general classical probability theory, but rather, as a specific class
of probabilistic models—a probabilistic theory—defined within a much broader
framework, in which individual models are defined by essentially arbitrary sets
of basic measurements and states, with the latter assigning probabilities (in an
elementary sense) to outcomes of the former. It is this framework that, if anything,
counts as a “non-classical” probability theory—but in fact, I think we should regard,
and refer to, this general framework, simply as probability theory, without any
adjective. Since it outruns what’s usually studied under that heading, I settle here

1 Given this, Gleason’s Theorem, as extended by Christensen and Yeadon, identifies the system’s
states with states (positive, normalized linear functionals) on that same von Neumann algebra,
and Wigner’s Theorem then identifies the possible dynamics with one-parameter groups of unitary
elements of that algebra. For the details, see Varadarajan (1985).
27 Dynamical States and the Conventionality of (Non-) Classicality 587

for the term general probability theory.2 Within this framework, in addition to QM,
we can identify classical probability theory, in the strict Kolmogorovian sense, as a
special, and, I will argue, equally contingent, probabilistic theory.
However—and this is my second point—even as applied to individual proba-
bilistic models, or, indeed, individual probabilistic theories, the distinction between
“classical” and “non-classical” is not so clear cut. One standard answer to the
question of what makes the probabilistic apparatus of QM “non-classical” is that
its state space (the set of density operators on a Hilbert space, or the set of states
on a non-commutative von Neumann algebra, depending on how general we want
to be) is not a simplex: the same state can be realized as a mixture—a convex
combination—of pure states in many different ways. Another, related, answer is
that, whereas in classical probability theory any two measurements (or, if you prefer,
experiments) can effectively be performed jointly, in QM this is not the case.
Neither of these answers is really satisfactory. In the first place, it is perfectly
within the scope of standard classical probability theory to consider restricted
sets of probability measures and random variables on a given measurable space:
the former need not be simplices, and the latter need not be closed under the
formation of joint random variables. We might also wish to allow “unsharp”
random variables having some intrinsic randomness (represented, depending on
one’s taste, by Markov kernels or by response-function valued measures). We might
try to identify as “broadly classical” just those probabilistic models arising in this
way, from a measurable space plus some selection of probability measures and
(possibly unsharp) random variables. However, once we open the door this far,
it’s difficult to keep it on its hinges. This is because there exist straightforward,
well known, and in a sense even canonical, recipes for representing essentially any
“non-classical” probabilistic model in terms of such a broadly classical one, either
by treating the state space as a suitable quotient of a simplex (in the manner of
the Beltrametti-Bugajski representation of quantum mechanics), or by introducing
contextual “hidden variables”. In fact, for systems with a high degree of symmetry,
including finite-dimensional quantum systems, there is another, less well known,
but equally straightforward classical representation: the probabilistic structure is
simply that of a single reference observable, and hence, entirely classical; states,
however, rather than corresponding simply to probability measures, are represented
by probability measure-valued functions on the system’s symmetry group, and
thus, as fundamentally dynamical objects. Conversely, a classical probability space
equipped with a family of suitably covariant such “dynamical states”—what I call a
dynamical classical model—can be interpreted as a generalized probabilistic model
in a canonical way.
At the level of individual probabilistic models, then, it seems that the distinction
between classical and non-classical is, at least mathematically, more or less one
of convention (and, pragmatically, of convenience). There is, of course, another

2 Theframework in question is broadly equivalent to that of “generalized probabilistic theories”.

Luckily, both phrases can go by the acronym GPT.
588 A. Wilce

consideration in play: that of locality. For probabilistic theories involving a notion

of composite systems that admits entangled states—including, of course, quantum
theory—the classical representations alluded to above are all manifestly non-local in
one sense or another. One reading of Bell’s Theorem is that, for such probabilistic
theories, there is a necessary tension between “classicality” and locality. Indeed,
one standard response to the existence of various classical representations is to
dismiss them precisely because they are non-local. But locality is very much a
physical constraint, rather than a “law of thought”. Thus, we should be careful not to
regard the probabilistic apparatus of quantum mechanics as a successor to classical
probability theory as a general account of reasoning in the face of uncertainty.
Rather, it is a particular, contingent, probabilistic physical theory, expressible in
terms of a much broader generalized probability theory—a framework that we may
or may not prefer to regard as non-classical, but which is in any event at most a very
conservative generalization of classical probability theory. It is a collateral aim of
this chapter to argue for this expansive view of what probability theory is.
A brief outline of this chapter is as follows. In Sect. 27.2, I give a condensed
introduction to the framework—or rather, one particular version of the framework—
of general(ized) probability theory, along the lines of Barnum and Wilce (2016).
It is another collateral aim of this paper to advertise this framework as offering
some additional measure of clarity to discussions of foundational issues in quantum
theory. In Sect. 27.3, I ask how one ought to characterize classical probabilistic
models in this framework, and review several more or less standard, or even
canonical, ways in which an arbitrary probabilistic model can be interpreted in terms
of a classical one. There is little here that is new, except possibly some streamlining
of the material. In Sect. 27.4, I introduce the dynamical classical models alluded
to above. In Sect. 27.5, I very briefly discuss some issues of locality, entanglement
and so forth in this context, and in Sect. 27.6, I gather a few concluding thoughts.
I have placed some technical material in a series of appendices. Appendix A
establishes a folk-result, to the effect that a probabilistic model in which any two
measurements are compatible in the sense of admitting a common refinement, is
essentially classical. Appendix B gathers some straightforward observations about
what are called semiclassical test spaces. Appendix C concerns a construction
that can be used to generate large numbers of highly symmetric probabilistic
models, or, equivalently, dynamical classical models. This follows, but in some
respects generalizes, material in Wilce (2009). To keep measure-theoretic details
to a minimum, I focus mainly (though not exclusively) on probabilistic models in
which there is a finite upper bound on the number of distinct outcomes in a basic
measurement.
Terminology: In the balance of this chapter, I will use the adjective classical
sometimes in a broad and informal way, sometimes in a more technical sense, and
sometimes ironically. I hope that context, occasionally aided by scare-quotes, will
help make my meaning clear in each case. The term classical representation is
also to be understood broadly and somewhat informally, to mean any mathematical
representation of one probabilistic model or theory in terms of another such
model or theory that is, in some reasonable sense, classical. At two points, my
27 Dynamical States and the Conventionality of (Non-) Classicality 589

mathematical usage may be slightly nonstandard: first, by a measurable space, I will

always mean a pair (S, ) where S is a set and  is an algebra, but not necessarily
a σ -algebra, of subsets of S; and by a measure on S, unless otherwise specified I
mean a finitely-additive measure. Secondly, by a discrete Markov kernel I mean a
function
 p : S × T → [0, 1], where S and T are sets, possibly infinite, such that
y∈T p(x, y) = 1 for every x ∈ S, where the sum is meant in the unordered sense.
Otherwise, my usage will be fairly standard. Terms likely to be unfamiliar to most
readers are defined when introduced.

27.2 Probability Theory vs Probabilistic Theories

This section provides a brief but self-contained tutorial on the mathematical

framework that I favor for a general probability theory. This is a slight elaboration
of a formalism developed by D. J. Foulis and C. H. Randall; see, e.g., Foulis and
Randall (1978, 1981). A more complete survey can be found in Barnum and Wilce
(2016).

27.2.1 Test Spaces, Probability Weights, and Probabilistic

Models

In many introductory treatments of probability theory, a probabilistic model is

defined as a pair (E, p) where E is the outcome-set of some experiment, and p
is a probability weight on E, meant to reflect the actual statistical behaviour of
the experiment. An obvious generalization is to allow p to vary, i.e., to consider
pairs (E, ) where  is some set of possible probability weights. For example, if
E = {0, 1}n , we might want to restrict attention to binomial probability weights
with various probabilities of success.
An only slightly less obvious generalization is to allow E, too, to vary:
Definition 2.1 (Test spaces) A test space is a set M of non-empty sets E, F, . . . .,
understood
 as the outcome-sets of various experiments or, as we’ll say, tests. If
X := M is the set of all outcomes of all tests in M, a probability weight on M
is a function α : X → [0, 1] such that x∈E α(x) = 1 for every E ∈ M.
I will write Pr(M) for the set of all probability weights on a test space M.
Note that Pr(M) is a convex subset of RX : any weighted average pα + (1 − p)β
of probability weights α, β ∈ Pr(M) is again a probability weight on M.
Definition 2.2 (Probabilistic models) A probabilistic model is a pair A =
(M, ) where M is a test space and  ⊆ Pr(M) is a convex set of probability
weights on M, which we call the states of the model. Extreme points of  are
called pure states.
590 A. Wilce

It will be convenient to denote a probabilistic model by a single letter A, B, etc.

When doing so, I write, e.g., A = (M(A), (A)). The assumption that (A) is
convex is meant to reflect the possibility of randomizing the preparation of states;
that is, if we have some procedure that produces state α and another that produces
state β, we could flip a suitably biased coin to product the state pα + (1 − p)β for
any 0 ≤ p ≤ 1. It is also common to equip a probabilistic model A with a preferred
group G(A) of “physical transformations” (Hardy 2001; Dakič and Brukner 2011;
Masanes and Müller 2011), and I will also do so in Sect. 27.4. Until then, however,
this extra structure will not be necessary. Finally, I will suppose in what follows
that  contains sufficiently many probability weights to separate points of X and
that M contains sufficiently many tests (and hence X, sufficiently many outcomes)
to distinguish the states in  (since otherwise, one would presumably just identify
outcomes, respectively states, that are not distinguishable). Another condition that I
will often (but not always) impose is that, for every outcome x ∈ X(A), there exists
at least one probability weight α ∈ (A) with α(x) = 1. A model satisfying this
condition is said to be unital.
Notice that we permit tests E, F ∈ M to intersect, that is, to share outcomes.
From a certain point of view, this is absurd: when we perform a measurement,
we usually retain a record of what measurement we’ve made! Nevertheless, we
may have good reasons to want to identify outcomes from distinct tests, and to
demand that physically meaningful operations on outcomes—symmetries, proba-
bility assignments, etc.—respect these identifications. The usual term for this is
non-contextuality. In most of the probabilistic models one encounters in practice,
such outcome-identifications are in fact made, and in any event, it is mathematically
more general to allow tests to overlap than to forbid them to do so. Moreover,
contextual probability weights can easily be accommodated by way of the following
construction.
Definition 2.3 (Semi-classical test spaces) A test space M is semi-classical iff
E ∩ F = ∅ for all distinct tests E, F ∈ M. The semi-classical cover of an arbitrary
test space M is the test space

 = {E
M  | E ∈ M },

 = {(x, E)|x ∈ E}.

where, for E ∈ M, E
The outcome-set of M  is thus X
 = {(x, E)|x ∈ E ∈ M}, i.e., the coproduct
of the sets E ∈ M. “Contextual” probability weights on M are best regarded
 In general Pr(M)
as probability weights on M.  will be very much larger than
Pr(M); indeed, if M contains infinitely many tests, Pr(M)  will be infinite
dimensional, even if Pr(M) is finite-dimensional.

 α on M is dispersion-free or deterministic iff

Definition 2.4 A probability weight
α(x) ∈ {0, 1} for every x ∈ X = M.
27 Dynamical States and the Conventionality of (Non-) Classicality 591

One of the most obvious distinctions between the probabilistic models associated
with “classical” systems and those associated with quantum-mechanical systems
is that the former have an abundance of dispersion-free states, while the latter,
 always has
absent superselection rules, have none at all. Notice, however, that M
an abundance of dispersion-free probability weights (just choose an outcome from
every test).
Definition 2.5 (Locally finite test spaces) A test space M is locally finite iff every
test E ∈ M is a finite set. Every test space is associated with a locally finite test
space M# , consisting of finite partitions of tests in M. That is, writing D(E) for
the set of finite partitions of a set E,

M# = D(E).
E∈M

If M is a test space, an event for M is a subset of a test (that is, an event in the
 to that test). We write Ev(M) for the set of all events of
standard sense pertaining
M, so that Ev(M) = E∈M P(E). Thus, the outcome-space for M# is precisely
the set Ev(M)  {∅} of non-empty events.
If α is a probability weight
 on M, then we define the probability of an event
a ∈ Ev(M) by α(a) = x∈a α(x). This evidently defines a probability weight on
M# . However, unless M is locally finite to begin with, M# will generally support
many additional “singular” probability weights. (Indeed, if M is semiclassical, one
can simply choose a non-principal ultrafilter on each E ∈ M.)
Lemma 2.6 Let M be locally finite with outcome-set X. Then the convex set
Pr(M) of probability weights on M is compact as a subset of [0, 1]X (with the
product topology).
Proof Since each E ∈ M(A) is finte, the pointwise limit of a net of probability
weights will continue to sum to unity over E. Thus, Pr(M) is closed, and hence,
compact, in [0, 1]X . *
)
It follows that if A is a probabilistic model with M(A) locally finite, the closure
of (A) in Pr(M(A)) will also be compact. Because it is both mathematically
convenient and operationally reasonable to do so, I shall assume from now on that,
unless otherwise indicated, all probabilistic models A under discussion have a
locally finite test space M(A) and a closed, hence compact, state space (A).
While I will make no direct use of it, it’s worth noting that the Krein-Milman
Theorem now tells us that every state in (A) is a limit of convex combinations
of pure states.
It is often useful to consider sub-normalized
 positive weights
 on a test space
M, that is, functions α : X := M → R+ such that x∈E α(x) =: r ≤ 1,
independently of E. In fact, these are simply probability weights on a slightly larger
test space:
592 A. Wilce

Definition 2.7 (Adjoining a null-outcome) If M is a test space with outcome-

space X, let ∗ be symbol not belonging to X, and define

M+ = {E ∪ {∗}|E ∈ M}.

We can think of ∗ as a “null outcome” representing the failure of a test, in a setting

in which we regard this failure as meaning the same thing regardless of which test is
performed. As an extreme example, ∗ might represent the destruction of the system
under study. It should be clear that a probability weight on M+ corresponds exactly
to a sub-normalized weight on M.

27.2.2 Some Examples

Example 2.8 (Kolmogorovian models) We can treat standard measure-theoretic

probability theory within the present framework in the following way. Let (S, ) be
a measurable space and let D(S, ) be the set offinite partitions of S by elements
of , regarded as a test space. Note that X = D(S, ) is simply   {∅}. A
probability weight on D(S, ), then, is a function p :   {∅} → [0, 1] summing
to unity on every finite partition of S by members of . It is straightforward to show
that such a function, extended to all of  by setting p(∅) = 0, is a finitely additive
probability measure on , and conversely. We can now construct a probabilistic
model (D(S, ), ) in various ways, for instance, by taking  to consist of all
such measures, or of all σ -additive probability measures, or of probability measures
absolutely continuous with respect to some given measure, or of a single probability
measure of interest, etc. I will refer to all models of the form (D(S, ), ), where
 is any convex set of probability measures on (S, ), as Kolmogorovian.3
Note that every point s ∈ S defines a dispersion-free probability weight on
D(S, ), namely, the point-mass δ s (b) = 1 if b ∈  with s ∈ b, and δ s (b) = 0
otherwise. (More generally, a dispersion-free probability weight on D(S, ) is an
ultrafilter on .)
Example 2.9 (Quantum models) If H is a Hilbert space, let X(H) denote H’s unit
sphere and M(H), the set of maximal pairwise orthogonal subsets of X(H), that
is, the set of unordered orthonormal bases of H. We shall call this the quantum test
space associated with H. Any density operator ρ on H gives rise to a probability
weight on M(H), which I also denote by ρ, namely ρ(x) = ρx, x.4 Let (H)
be the set of all probability weights on M(H) arising from density operators in this
way: then A(H) := (M(H), (H)) is a probabilistic model representing the same

3 This is one setting in which we would often not require  to be closed. For instance, if  =

σ (S, ), then  is not closed in the topology of event-wise convergence.

4 Gleason’s Theorem tells us that if dim(H) > 2, all probability weights arise in this way, but we

will not make use of this fact.

27 Dynamical States and the Conventionality of (Non-) Classicality 593

quantum-mechanical system one associates with H, but in a manner that puts its
probabilistic structure in the foreground. Note that only if H is finite-dimensional,
is (H) compact.
While the model above is mathematically attractive, it is more usual to identify
measurement outcomes with projection operators. The projective quantum test
space consists of sets of projections pi on H summing to 1. If ρ is a density operator
on H, we obtain a probability weight (also denoted by ρ) given by ρ(p) = Tr(ρp).
Letting  consist of all such pobability weights, we obtain what we can call the
projective quantum model. More generally still, if A is a von Neumann algebra (or,
any ring with identity), the collection M(A) of sets E of projections
for that matter,
p ∈ A with p∈E p = 1 is a test space, on which every state on A gives rise to
a probability weight by evaluation. In this case, we might identify  with the state
space (in the usual sense) of A, which is weak-∗ compact, but which contains a host
of non-normal states.
Two pathological examples The following simple examples are useful as
illustrations of the range of possibilities comprended by the formalism sketched
above.
Example 2.10 (Two bits) Let M = {{a, a  }, {b, b }} be a test space consisting of
two yes-no tests, one with outcomes a and a  , the other with outcomes b and b .
The space of all probability weights on M is essentially the unit square, since such
a weight α is determined by the pair (α(a), α(b)). While the state space is not a
simplex, pure states are dispersion-free. The square bit B and diamond bit B  are
the probabilistic models having the same test space, namely

M(B) = M(B  ) = M,

but the two different state spaces pictured below in Fig. 27.1.
The square bit figures, directly or indirectly, in the large literature on “PR boxes”
inaugurated by Popescu and Rohrlich (1994). While the state space of B  is affinely
isomorphic to that of B, it interacts with M differently. While both models are
unital, in B  , for each outcome x there is a unique state δ x with δ x (x) = 1.

Fig. 27.1 Two bits

594 A. Wilce

Fig. 27.2 A “Greechie

diagram” of the Wright
c
Triangle, in which outcomes
belonging to a commont test
lie along a smooth curve
(here, a straight line)

z y

a x b

Examples 2.11 (The Wright triangle5 ) A more interesting example consists of three
overlapping, three-outcome tests pasted together in a loop, as in Fig. 27.2:

Pr(M) = {{a, x, b}, {b, y, c}, {c, z, a}}

The space Pr(M) of all probability weights on M is three-dimensional, since

such a weight is uniquely determined by the triple (α(a), α(b), α(c)). It is not
difficult to check that is spanned by four dispersion-free states, determined by

α(a) = α(y) = 1, β(b) = β(z) = 1, γ (c) = γ (x) = 1 and δ(x) = δ(y) = δ(y) = 1,

and a fifth, non-dispersion free pure state

(a) = (b) = (c) = 1/2, (x) = (y) = (z) = 0.

In particular,  is not a simplex, and not all pure states are dispersion-free.

27.2.3 Probabilistic Models Linearized

It is often convenient to consider the vector space V(A) ≤ RX(A) spanned by the
state space (A). This carries an obvious pointwise order. Writing V(A)+ for the
cone of all α ∈ V(A) with α(x) ≥ 0 for all x ∈ X(A), it is not difficult to see that

5 So well known in the quantum-logic community as to be almost a cliche, this gives rise to the
simplest quantum logic that is an orthoalgebra but not an orthomodular poset. See Wilce (2017)
for details. The example is named for R. Wright, one of Randall and Foulis’s students.
27 Dynamical States and the Conventionality of (Non-) Classicality 595

each α ∈ V(A)+ has the form α = rβ where β ∈ (A) and r ≥ 0. Moreover, the
coefficient r is unique, so (A) is a base for the cone V(A)+ .
An effect on A is a functional f : V(A) → R with 0 ≤ f (α) ≤ 1 for α ∈ (A).
Every outcome x ∈ X(A) determines an effect  x ∈ V(A)∗ by evaluation, that is, by

x (α) = α(x) for all α ∈ V(A). We can regard an arbitrary f as representing an “in-
principle” measurement outcome, not necessarily included among the outcomes in
X(A), but consistent with the convex structure of (A): f (α) is the probability
of the effect f occurring (if measured) in the state α. There is a unique unit
effect uA given by u(α) = 1 for all α ∈ (A), representing the trivial outcome
“something happened”. If f is an effect, so is uA − f =: f  ; we interpret f 
as representing the event that f did not occur. Thus, the pair {f, f  } is an in-
principle test (measurement, experiment) not necessarily
 included in M(A). More
generally, if f1 , . . . , fn is a set of effects with i fi = uA , we regard {f1 , . . . , fn }
as an “in principle” measurement or experiment—the standard term is “discrete
observable”—not necessarily included in the catalogue M(A), but consistent with
the structure of (A). Each test E ∈ M(A) gives rise to a discrete observable
{
x |x ∈ E} in this sense.
Examples 2.12
(a) As an illustration, let (S, ) be a measurable space, and let  = (S, )
be the space of finitely additive probability measures on . Then V() is the
space of all signed measures μ = μ+ − μ− on , where μ+ and μ− are
positive measures. Any measurable
- function ζ : S → [0, 1] defines an effect
a through the recipe a(μ) = S ζ (s)dμ(s). It is common to think of such an
effect as a “fuzzy” or “unsharp” version of an indicator function (modelling,
perhaps, a detector that responds with probability ζ (s) when the system is in
a state s). The unit effect is the constant function 1. A discrete observable,
accordingly, represents a generalization of a “fuzzy” or unsharp version of a
finite measurable partition.
(b) In the case of a quantum-mechanical model, say A = A(H), if we identify
(A) with the set of density operators on H, then V(A) can be identified with
the space of all compact self-adjoint operators on H, ordered in the usual way;
V(A)∗ is then naturally isomorphic, as an ordered vector space, to the space
of all bounded self-adjoint operators on H, and an effect is an effect in the
usual quantum-mechanical sense, that is, a positive self-adjoint operator a with
0 ≤ a ≤ 1, where 1 is the identity operator on H, which serves as the unit effect
in this case. A discrete observable corresponds to a discrete positive-operator
valued measure (POVM) with values in N.
Unlike a test in M(A), a general discrete observable on A can have repeated
values. Nevertheless, by considering their graphs, such observables can be organized
into a test space, as follows. Here, [0, uA ] denotes the set of all effects on A, and
(0, uA ] denotes the set of non-zero effects.
Definition 2.13 (Ordered tests) The space of ordered tests over the convex set 
is the test space Mo () with
596 A. Wilce

(a) Outcome-space Xo () := N × (0, u];

 those finite sets a ⊆ X that are graphs of mappings a : I → (0, u] with
(b) Tests, o

i∈I ai = u, where I ⊆ N is finite.

In other words, tests in Mo () are (graphs of) finite discrete observables with
finite value-spaces I ⊆ N. Every state α ∈  defines a probability weight α o on
Mo () by α o (i, a) = α(a).6 In effect, Mo () is the largest test space supporting
 as a separating set of probability weights.

27.2.4 Probabilistic Theories

A probabilistic theory ought to be more than a collection of probabilistic models.

One also needs a collection of allowed mappings between such models turning
C into a category. There are many plausible ways to define a morphism between
probabilistic models, some more restrictive, some very general (see, e.g., Foulis and
Randall 1978; Barnum and Wilce 2016). For the purposes of this chapter, it will
be sufficient to consider the following, relatively strict definition (due to Foulis and
Randall 1978). Recall that Ev(M) denotes the set of events (subsets of tests) for
a test space M. Given two events a, b ∈ Ev(M), we write a ⊥ b to mean that
a ∩ b = ∅ and a ∪ b is again an event. That is, a ⊥ b iff a and b are compatible
(jointly testable), and mutually exclusive.
Definition 2.14 (Interpretations) Let M and M be test spaces with outcome-
sets X and X . An interpretation from M to M is a mapping
φ : X −→ Ev(M

)
 
such that if x ⊥ y in X, then φ(x) ⊥ φ(y) in Ev(M ) and x∈E φ(x) ∈ M for
all E ∈ M.
In other words, an interpretation φ : M → M allows us to regard (to interpret!)
each test E ∈ M as a coarse-grained version of a test in M , in a non-contextual
way.
Definition 2.14 (Continued) The following terminology will be useful. An inter-
pretation φ : M → M is outcome-preserving iff φ(x) is either empty or a
singleton for every x ∈ X, in which case we can identify
 φ with the corresponding
mapping from the outcome space X = M to the outcome space
partial 
X = M . If φ(x) is non-empty for every x ∈ M, then φ is positive. A
positive, outcome-preserving interpretation φ : M → M , amounts to a mapping
φ : X → X with φ(M) ⊆ M . Where this mapping φ is injective, allowing us to
identify M with a subset of M , we call it an embedding. Where φ(M) = M ,

6 Itcan be shown that, conversely, every probability weight on Mo () is determined by a finitely-
additive normalized measure on the effect algebra [0, u], and hence, by a positive linear functional
in V(A)∗∗ . Hence, probability weights on Mo () that are continuous on (0, u] in the latter’s
relative weak-∗ topology, arise from elements of .
27 Dynamical States and the Conventionality of (Non-) Classicality 597

we will say that φ is a cover of M by M. Where φ is bijective on both outcomes

and tests, we call it an isomorphism of test spaces.
Examples 2.15
(a) If (S, ) and (T , &) are measurable spaces, any measurable mapping f : S →
T gives rise to an outcome-preserving interpretation φ : D(T , &) → D(S, ),
given by φ(b) = {f −1 (b)} when f −1 (b) = ∅, and φ(b) = ∅ otherwise.
(b) Suppose K1 and K2 are compact convex sets and f : K1 → K2 is an affine
mapping, we can define an outcome-preserving interpretation φ : Mo (K2 ) →
Mo (K1 ) by setting

{(i, b ◦ f )} b ◦ f = 0
φ(i, b) =
∅ b◦f =0

(c) For an example of a positive but not outcome-preserving, interpretation,

consider the inclusion mapping Ev(M)  {∅} → Ev(M): we can understand
this as an interpretation M# → M taking a non-empty event qua outcome of
M# to the same event viewed as a set of outcomes of M.
Notice that any interpretation φ : M1 → M2 gives rise to an affine mapping
φ ∗ : Pr(M2 ) → Pr(M1 ), defined for β ∈ (M2 ) by φ ∗ (β) = β ◦ φ.
Definition 2.16 (Interpretations of models) An interpretation φ : A → B from a
probabilistic model A to a probabilistic model B is an interpretation φ : M(A) →
M(B) such that φ ∗ ((B)) ⊆ (A). We say that φ is
(a) an embedding iff φ is an embedding of test spaces and φ ∗ is surjective;
(b) a cover iff φ is a cover of test spaces, in which case φ ∗ is injective.
The affine mapping φ ∗ : (B) → (A) extends uniquely to a positive linear
mapping φ ∗ : V(B) → V(A), satisfying uA ◦ φ ∗ = uB . An apparently much
more general notion of a morphism from a model A to a model B is a positive
linear mapping  : V(B) → V(A) with the feature that uA ◦  ≤ uB , that is,
uB ((β)) ≤ 1 for all β ∈ (B) (Barnum and Wilce 2016). Such a map represents
a possibly “lossy” process that, given as input a state β ∈ (B), produces an output
state α := (β)/uB ((β)) with probability uB ((β))—or, if uB ((β)) = 0,
simply destroys the system. However, given such a mapping , and letting ∇(A)
denote the convex set of sub-normalized states rα were r ∈ [0, 1] and β ∈ (B),
we can define an interpretation φ : Mo (∇(A)) → Mo ((B)) as in Example 2.15
(b) above. Thus, we actually lose very little, if any, generality in restricting attention
to categories of probabilistic models and interpretations. In the remainder of this
chapter, a probabilistic theory will always mean such a category.
598 A. Wilce

27.3 Classicality and Classical Representations

There are various ways of representing probabilistic models in terms of classical

probabilistic models, most of them reasonably familiar in the context of QM. In
this section, I want to discuss two of these that are rather canonical, both in the
informal sense that they are more or less obvious, and apply to more or less arbitrary
probabilistic models, and also in the technical sense that the constructions involved
are actually functorial. Of these, it is my sense that only the first is really well known.

27.3.1 Classical Models and Classical Embeddings

Before we can discuss these representations, we need to address head on the central
concern of this chapter: what does it mean (what ought it to mean) for a probabilistic
model A to be “classical”? Whatever our final answer, presumably we would at least
wish to consider models A(S, ) := (D(S, ), σ (S, )) as classical. However,
as discussed earlier, even classically it won’t do to restrict attention to such models:
any model of the form (D(S, ), ) where  is any subset of σ (S, ), should
almost certainly count as “classical”, in the sense of falling within the scope of
classical probability theory (Lindsay 1995).
We should also pause to ask whether or not the restriction to σ -additive measures
should be regarded as one of the defining features of classicality. If μ is a finitely
additive measure on an algebra  of sets, then letting T = β() be the Stone space
of , regarded as a Boolean algebra, one finds that, owing to the fact that elements
of  are clopen in T and T is compact, no countable disjoint family of sets in  has
union equal to T . Hence, every measure on (T , ) is countably additive by default.
In different language, whether or not a measure on a Boolean algebra  is countably
additive depends on the particular representation of  as an algebra of sets.
For this reason, it seems prudent to allow any Kolmogorovian model, in the sense
of Example 2.8—that is, any probabilistic model of the form (D(S, ), ), where
 is any algebra of subsets of S and  is any convex set of probability measures
on , countably additive or otherwise, to count as “classical”. But if this is so, it
seems equally reasonable to admit as classical models of the form (M, ) as long
as M ⊆ D(S, ) and  ⊆ (S, ). That is, models admitting embeddings, in the
sense of Definition 2.16, into Kolmogorovian models in the sense of Example 2.8,
should themselves be considered “classical” in a broad sense.
When does an abstract model (M, ) admit such a classical embedding? If
φ : M(A) → D(S, ) is an interpretation, then every point-mass δ s in (S, )
pulls back along φ to a dispersion-free probability weight on M(A). If φ is to be
an embedding of test spaces, there must be enough of these dispersion-free states to
separate outcomes in X(A). Moreover, if φ is to be an embedding of probabilistic
models, every state α ∈ (A) must belong to the closed convex hull of these
dispersion-free weights.
27 Dynamical States and the Conventionality of (Non-) Classicality 599

Definition 3.1 (UDF models) A probabilistic model A is unitally dispersion-free

or UDF iff every state α ∈ (A) lies in the closed convex hull of the set of
dispersion-free probability weights on M(A).
Versions of the following can be found in many places in the quantum-logical
literature.
Lemma 3.2 A model admits a classical embedding iff it is UDF.
Proof The “only if” direction is clear from the discussion above. For the converse,
suppose (M, ) is UDF, and  let S be the set of dispersion-free probability weights
on M. For each x ∈ X := M, let

ax := {s ∈ S|s(x) = 1} ⊆ S.

Then {ax |x ∈ E} is a partition of S. Let  be the algebra of subsets of S generated

by sets of the form ax . Then the mapping φ : X →  given by φ(x) = ax gives rise
to an outcome-preserving interpretation M → D(S, ), which is an embedding
if and only if S separates points of X. The mapping φ ∗ : R → RX given by
φ ∗ (μ)(x) = μ(ax ) is continuous with respect to the product topologies on these
spaces, and for each s ∈ S, φ ∗ (δ s ) = s. It follows that con(S) = φ ∗ ((S, )).
Hence, the probability weights in con(S) have a common classical explanation.
Given any closed convex set  ⊆ con(S), we then have an embedding of the model
(M, ) into the Kolmogorovian model (D(S, ), (φ ∗ )−1 ()). *
)
Let me stress that it is not sufficient for M(A) simply to have a large supply
of dispersion-free probability weights: we also require that all of the allowed states
α ∈ (A) arise as (limits of) averages of these. As an example, the Wright Triangle
admits many dispersion-free states, but also a non-dispersion-free extreme state; this
state can not be explained in terms of a classical embedding.
Of course, if M(A) has no dispersion-free states at all, then such a classical
embedding is impossible in any case. Gleason’s Theorem tells us that (H) has no
dispersion-free states if dim(H) ≥ 3; this is the substance of one of the strongest
“no-go” theorems for hidden variables. However, this is far from the end of the story.
For one thing, the classical embeddings considered above are non-contextual. That
is, if x ∈ E ∩ F , where E, F ∈ M(A), then φ(x) does not depend on whether we
regard x as an outcome of E or of F . As was realized very early on making φ(x)
depend also on the choice of “measurement context” (e.g., E or F , above) allows
for much more flexibility.
Besides the existence of an abundance of dispersion-free states, a distinctive
feature of Kolmogorovian models is that any two partitions E, F ∈ D(S, ) can be
regarded as coarse-grained versions of a common refinement: if

G = { a ∩ b | a ∈ E, b ∈ F, a ∩ b = ∅ }

then G ∈ D(S, ), and every a ∈ E, and likewise every b ∈ F , is in an obvious

sense equivalent to a subset of G. This allows us to make a simultaneous joint
600 A. Wilce

measurement of E and F , by performing a measurement of G. One can define a

general notion of the refinement of one test by another in the context of an arbitrary
probabilistic model, as follows:
Definition 3.3 (Refinement) A test E ∈ M(A) refines a test F ∈ M(A) iff there
exists a surjection f : E → F such that α(y) = α(f −1 (y)) for every y ∈ F and
every state α ∈ (A).
Recall that a probabilistic model A is unital if every outcome x ∈ X(A) has
probability one in at least one state, i.e., there is at least one state α ∈ (A) with
α(x) = 1.
Theorem 3.4 Let A be a unital model in which every pair of tests has a common
refinement. Then A has a classical embedding.
The proof is given in Appendix A.
Conversely, of course, if A has a classical embedding, then M(A) embeds
in a test space D(S, ), in which every pair of tests has a common refinement.
Thus, a necessary and sufficient condition for a unital model A to have a classical
embedding is that it be possible to embed A in a model B in which every pair of
tests has a common refinement. To this extent, one might reasonably say that it is the
existence (or at least, the mathematical possibility) of joint measurements that is the
hallmark of “classical” probabilistic theories. However, it is very important to note
again at this point, that this is an entirely contingent feature of classical probability
theory. That is, there is no “law of thought” that tells us that any two measurements
or experiments can be performed jointly.

27.3.2 Classical Extensions

An embedding is not the only way of explaining one mathematical object in terms
of another. One can also consider representing the object to be explained as a
quotient of a more familiar object. In connection with probabilistic models, this
line of thinking was explored by Holevo (1982) and, slightly later, by Bugajsky
and Beltrametti (Beltrametti and Bugajski 1995; see also Hellwig and Singer 1991;
Hellwig et al. 1993.) In the literature, this sort of representation of a quantum model
is usually called a classical extension. The basic idea is that any convex set can arise
as a quotient of another under an affine surjection. As a simple example, a square
can arise as the projection of a regular tetrahedron (a 4-simplex) on a plane.
To develop this idea further, we will need a bit of background on Choquet theory
(Alfsen 1971). Recall that the σ -field of Baire sets of a compact Hausdorff space
S is the smallest σ -algebra containing all sets of the form f −1 (0) where f ranges
over continuous real-valued functions on S. Let K be a compact convex set, that
is, a compact, convex subset of a locally convex topological vector space V , and
let o (K) be the set of Baire probability measures thereon. This is an infinite-
dimensional simplex in the sense that V(o (K)) is a vector lattice (Alfsen 1971).
For each μ ∈ (K), define the barycenter of μ,  μ ∈ V ∗∗ , by
27 Dynamical States and the Conventionality of (Non-) Classicality 601



μ(f ) = f dμ
K

for all functionals f ∈ V ∗ . Using the Hahn-Banach Theorem, one can show that

μ ∈ K (Alfsen 1971, I.2.1). Thus, we have an affine mapping μ →  μ from o (K)
to K. The barycenter of the point-mass δ α at α ∈ K is evidently α, so this mapping
is surjective. In this sense, every compact convex set is the image of a simplex under
an affine mapping.
Definition 3.5 (Classical extensions) A classical extension of a probabilistic
model A consists of a measurable space (S, ) and an affine surjection
q : σ (S, ) → (A).
This makes no reference to M(A). However, the affine surjection q : o (S) →
(A) can be dualized to yield a mapping q ∗ : X(A) → Aff(o (S)) taking every
outcome x ∈ X(A) to the functional q ∗ (x)(μ) =  μ(x), which is evidently an
effect on (S). Now, such an effect can be understood as a “unsharp” version
 of an
indicator function, as discussed in Example 2.12(a) Accordingly, since x∈E q ∗ (x)
is identically 1, q ∗ is an effect-valued weight on M(A), representing each test as
a partition of 1 by such fuzzy indicator functions, that is, q ∗ (E) is a (discrete)
“unsharp random variable” on S. For more on this, see Bacciagaluppi (2004).
In such a representation, in other words, the state space is essentially classical
(it’s simply a set of probability measures), while outcomes and tests become
“unsharp”. While this may represent a slight extension of the apparatus of standard,
Kolmogorovian probability theory, it is certainly within the scope of classical
probability theory in the somewhat wider sense that concerns us here.
Another way to put all this is that a classical extension comes along with
(and determines) an interpretation M(A) → Mo ((S, )), where the latter
is the space of ordered tests on (S, ), in the sense of Definition 2.13. This
then gives us an interpretation of models A → (Mo ((S, )), (S, )). In
this sense, a classical extension is an “unsharp” version of a classical embedding.
However, as the discussion above shows, every probabilistic model has a canonical
classical extension obtained by taking S = (A) and  the Baire field of .
This construction is even functorial, since the construction A → (A) is (the
object part of) a contravariant functor from the category of probabilistic models
to the category of compact convex sets, while K → 0 (K) and K → Mo (K)
are, respectively, a covariant endofunctor on the category of compact convex sets
and continuous affine maps, and from this category to the category of test spaces
and interpretations. Putting these together gives us a covariant functor A →
(Mo (0 ((A))), 0 ((A))) from arbitrary (locally finite) probabilistic models
to unsharp Kolmogorovian models.
One can also construct a classical extension in which the carrier space is the set
Kext of extreme points of K. This need not be a Borel subset of K; however, it
becomes a measurable space if we let ext be the trace of K’s Baire field o on
Kext , i.e., ext = { b ∩ Kext | b ∈  }. If μ is a probability measure on ext , then
we can pull this back along the boolean homomorphism φ : b → b ∩ Kext to a
602 A. Wilce

probability measure μ ◦ φ on o , the barycenter of which, as defined above, defines

a barycenter for μ; that is, we take  μ := μ  ◦ φ. The Bishop-deLeeuw Theorem
(see Alfsen 1971, I.4.14) asserts that every point of K can be represented as the
barycenter of a σ -additive probability measure on ext .
We can regard probability measures μ in the simplex (S), S = (A)ext , as
representing ensembles of pure states, that is, preparation procedures that produce
a particular range of pure states, with prescribed probabilities. The quotient map
(S) → (A) simply takes each such ensemble to its probability-weighted
average. Where (A) is not a simplex, many different ensembles will yield
this same average; operationally, that is, using the measurements available in
M(A), distinct ensembles averaging to the same state are indistinguishable, so
treating a state α ∈ (A) as “really” arising from one, rather than another, such
ensemble represents a kind of contextuality, what Spekkens (2005) calls preparation
contextuality.

27.3.3 Semiclassical Covers

 of a test space M is given

Recall from Sect. 27.2.1 that the semiclassical cover, M,
by

 = {E
M  | E ∈ M },

where, for E ∈ M, E  = {(x, E)|x ∈ E}. The outcome-set of M  is thus X =

{(x, E)|x ∈ E ∈ M}, i.e., the coproduct of the tests E ∈ M. There is an obvious
outcome-preserving interpretation π : M  → M, given by π (x, E) = {x}. Any
probability weight α on M pulls back along this to a probability weight  α := π ∗ (α)

on M, given by  α (x, E) := α(x) for all x ∈ E ∈ M(A). Given a model A, we
define the semiclassical cover of A to be the model A 
 := (M(A), (A)}, where


M(A) is the semiclassical cover of M(A) and  (A) := {  α | α ∈  }. The
mapping α →  (A) is isomorphic to (A). Thus, A
α is an affine injection, so  
differs from A only in the structure (that is, the comparative lack of structure) of its
test space.
Since M(A) is semiclassical, its space Pr(M)  of probability weights is
essentially the Cartesian product E∈M (E) of the finite-dimensional simplices
(E) of probability weights on the various tests E ∈ M. Indeed, we can represent
α ∈ Pr(M(A)) uniquely as (α E ) where α E is α’s restriction to E ∈ M(A).
Hence, M  supports a wealth of dispersion-free probability weights, namely, those
weights obtained by selecting a dispersion-free probability weight—a point-mass—
from each simplex (E).
In what follows, I’ll write S(A) for the set of all such dispersion-free states on

M(A). It is easy to show that these are exactly the extreme probability weights on

M(A), and that S(A) is closed in Pr(M(A))  (see Appendix B). This allows us
27 Dynamical States and the Conventionality of (Non-) Classicality 603

to represent every α ∈ Pr(M)  as the barycenter of a Borel probability measure

on S(A) (as every Baire probability measure on a compact space has a unique
Borel extension). Letting (S(A)) denote the simplex of Borel probability measures
on S(A), this gives us an affine surjection φ : (S(A)) → Pr(M(A)).  Also,
ψ : (x, E) → {α ∈ S|α(x, E) = 1} gives an interpretation of each test
 ∈ M(A)
E  as an element of D(S, ), where  is the Borel algebra of S(A),
and it is straightforward that φ = ψ ∗ . We therefore have the following picture:

D (S(A), ) (S(A))

ψ φ = ψ∗

M(A) (A)
M (A) (A))
Pr(M
π π∗

The dual mapping π ∗ (α)(x, E) = α(x) is injective, while ψ ∗ is surjective: again,

(A) is a convex subset of a quotient of a simplex.
However, in this representation observables associated with tests
E ∈ M(A) correspond to sharp classical random variables on S(A): letting

ax = {s ∈ S(A)|s(x, E) = 1},


{ax |x ∈ E} is a partition of S(A).
This gives us a very simple representation of an arbitrary model as a quotient of a
sub-model of Kolmogorovian model. To the extent that a representation of a model
A as a quotient of a model B allows us to “explain” the model A in terms of the
model B, and to the extent that representing a model B as a submodel of a model
C also allows us to explain B in terms of C, then—to the extent that “explains” is
transitive—we see that every model can be explained in terms of a classical one.
This construction is even functorial: if φ : A → B is an interpretation from a model
A to a model B, there is a canonical extension of φ to an interpretation φ:A → B,
given by φ(x, E) = (φ(x), φ(E)). Hence, an entire probabilistic theory can be
rendered essentially classical, if we are willing to embrace contextuality.
Hidden Variables This is as good a place as any to clarify how the foregoing
discussion connects to traditional notions of “hidden variables” (or “ontological
representations” (Spekkens 2005)). Briefly, a (not necessarily deterministic, not
necessarily contextual) hidden variables representation of a probabilistic model
A consists of a measurable space  and, for every test E ∈ M, a conditional
probability distribution p( · |λ,
E) on E—that is, p(x|λ, E) is a non-negative real
number for each x ∈ E, and x∈E p( x | λ, E ) = 1. It is also required that, for
every state α ∈ (A), there exist a probability measure μ on  such that for every
x ∈ X(A) and every test E with x ∈ E,

p(x|λ, E)dμ(λ) = α(x). (27.1)

604 A. Wilce

Now, if we rewrite p(x|λ, E) as pλ (x, E), we see that pλ is simply a probability


weight on M(A). We can then write (27.1) as
 
α(x) = pλ dμ(λ) (x, E) (27.2)


which merely asserts that α is the μ-weighted average of a family of probability


weights on M(A), parametrized by .

27.3.4 Discussion

In view of the classical representations discussed in Sects. 27.3.2 and 27.3.3, it

seems that we can always understand a “generalized probabilistic model”, or,
indeed, a generalized probabilistic theory, in terms of a classical model or theory.
The only departure from classical Kolmogorovian probability theory, in these
representations, is a restriction—which one can regard as epistemic or as “physical”,
as the case warrants—on which ensembles of probability measures we can prepare,
which (classical) measurements we can perform, and hence, on which measurement
outcomes and which mixtures of states we can distinguish. To be sure, we are under
no obligation to take ensembles in (ext ) or dispersion-free states on M  seriously
as part of any given physical theory, and we may also prefer to do without them for
reasons of mathematical economy; but both are mathematically meaningful, and
certainly have a place in any developed general probability theory. To this extent,
arbitrary probabilistic models, and even probabilistic theories, have fairly canonical
“explanations” in Kolmogorovian terms.
In the next section I will go just a step further and point out that, for models
with certain strong symmetry properties (including quantum models), there is
available a rather different kind of “classical” representation: one in which the
model’s probabilistic apparatus is entirely classical—there is, in fact, only one basic
measurement. In such a representation it is, rather, the interplay between a system’s
symmetries, its states, and this perfectly classical probabilistic apparatus, that is (in
a sense I try to pin down) perhaps not quite “classical”.

27.4 Dynamical Models and Dynamical States

One of the things that most conspicuously distinguishes quantum probabilistic

models from more general ones is the their very high degree of symmetry: given
any two pure quantum states, there is a unitary operator taking one to the other, and
likewise, given any two projective orthonormal bases, there is a unitary mapping
one to the other. In this section, I want to sketch a different kind of classical
27 Dynamical States and the Conventionality of (Non-) Classicality 605

representation, available for locally finite models whenever such a great deal of
symmetry is in play. This is mathematically similar to the representation in terms
of the semiclassical cover; conceptually, however, it is leaner, in that it privileges
a single observable, and encodes the structure of the model in the ways in which
probabilities on this observable can change.

27.4.1 Models with Symmetry

In this section, we will consider probabilistic models A that, in addition to the test
space M(A) and state space (A), are equipped with a distinguished symmetry
group G(A) acting on X(A), M(A) and (A). In most applications, G(A) will be
a lie group and, in our finite dimensional setting, compact.
Definition 4.1 (Symmetries
 of test spaces) A symmetry of a test space M with
outcome-set X = M is an isomorphism g : M → M, or, equivalently, a
bijection g : X → X such that ∀E ⊆ X, g(E) ∈ M iff E ∈ M. We write
Aut(M) for the group of all symmetries of M under composition.
Note that Aut(M) also has a natural right action on Pr(M), since if g ∈
Aut(M) and α ∈ Pr(M), then α ◦g is again a probability weight, and the mapping
g ∗ : α → α ◦ g is an affine bijection Pr(M) → Pr(M).
Definition 4.2 (Dynamical models) A dynamical probabilistic model is a proba-
bilistic model A with a distinguished dynamical group G(A) ≤ Aut(M(A)) under
which (A) is invariant (Barnum and Wilce 2016).
I use the terms dynamical model and dynamical group because in situations in
which G(A) is a Lie group, a dynamics for (or consistent with) the model will
be a choice of a continuous one-parameter group g ∈ Hom(R, G(A)), that is, a
mapping t → gt with gt+s = gt gs , which we interpret at tracking the system’s
evolution over time: if α is the state at some initial time, gt (α) = α ◦ g−t is the state
after the elapse of t units of time. That g is a homomorphism encodes a Markovian
assumption about the dynamics, namely, that a system’s later state depends only on
its initial state and the amount of time that has passed, rather than on the system’s
entire history.
Definition 4.3 (Symmetric and fully symmetric models) A dynamical proba-
bilistic model A is symmetric iff G(A) acts transitively on the set M(A) of tests,
and the stabilizer, G(A)E , of a (hence, any) test E ∈ M acts transitively on E. Note
that this implies that all tests have a common size. We say that A is fully symmetric
iff, additionally, for any test E ∈ M(A) and any permutation σ : E → E, there
exists at least one g ∈ G(A) with gx = σ (x) for every x ∈ E.
An equivalent way to express full symmetry is to say that all tests E, F ∈ M(A)
have the same cardinality, and, for every bijection f : E → F , there exists some
g ∈ G(A) with gx = f (x) for every x ∈ E.
606 A. Wilce

The quantum probabilistic model A(H) is fully symmetric under the unitary
group U (H), since any bijection between two (projective) orthonormal bases for H
extends to a unitary operator.
We can reconstruct a symmetric test space M(A) from any one of its tests, plus
information about its symmetries (Wilce 2009). Suppose we are given, in addition
to the group G(A), a single test E ∈ M(A), an outcome xo ∈ E, and the two
subgroups

H := { g ∈ G(A) | gE = E } and K := { g ∈ G(A) | gxo = xo }.

We then have a canonical bijection G(A)/K → X(A), sending gK to gxo , and

sending {hK|h ∈ H } bijectively onto E; M(A) can then be recovered as the orbit of
E in X(A). Thus, more abstractly, one could start with a triple (G, H, K) where G
is a group and H, K are subgroups of G, define X := G/K and E = {hK|h ∈ H },
and set M = {gE|g ∈ G}. If H acts as the full permutation group on E, M will be
fully symmetric under G. The choice of a closed convex set  ⊆ Pr(M), invariant
under G, completes the picture and gives us a G-symmetric probabilistic model.
The simplest cases are those in which H S(E). This holds for ordinary
quantum test spaces, since every permutation of an orthonormal basis determines
a unique unitary. In what follows, we concentrate on this case. (For a projective
quantum test space, in which E is a maximal set of pairwise orthogonal rank-one
projections, H is generated by permutation matrices acting on E, plus unitaries
commuting with these projections.)

27.4.2 A Representation in Terms of Dynamical States

We now reformulate these ideas in a way that gives more prominence to the chosen
test E ∈ M(A). In brief: if we hold the test E fixed, but retain control over both
the state space (A) and the dynamical group G(A), we obtain a mathematically
equivalent picture of the model, but one in which the probabilistic structure is, from
a certain point of view, essentially classical.
Consider the following situation: one has some laboratory apparatus, defining
an experiment with outcome-set E. This is somehow coupled to a physical system
having a state-space , governed by a Lie group G. That is, G acts on  in such
a way that all possible evolutions of this system are described by an initial state
α o ∈  and a one-parameter subgroup (gt )t∈R of G. It is harmless, and will be
convenient, to take  to be a right G-space, that is, to denote the image of state
α ∈  under the action of g ∈ G by αg. We may suppose that the way in which the
apparatus is coupled with the system manifests itself probabilistically, by a function

p :  × E → [0, 1]
27 Dynamical States and the Conventionality of (Non-) Classicality 607

giving the probability p(α, x) to obtain outcome x ∈ E when the system is in

state α.7 I will assume in what follows that the probability weights p(α, ·) separate
outcomes in E, that is, that for all outcomes x, y ∈ E,

( ∀α ∈  p(α, x) = p(α, y) ) ⇒ x = y. (27.3)

(If not, we can factor out the obvious equivalence relation on E.) I will also suppose
that the experiment E and the group G together separate states, that is, knowing
how the probabilities p(αg, x) vary with g ∈ G, for all outcomes x ∈ E, is enough
to determine the state α uniquely. In other words,

( ∀g ∈ G, ∀x ∈ E p(αg, x) = p(βg, x) ) ⇒ α = β (27.4)

Notice that for each state α ∈ , we can define a mapping  α : G → (E),

where (E) is the simplex of probability weights on E, given by


α (g)(x) = p(αg, x).

α as a mapping G × E → [0, 1] with x∈E 
Alternatively, we can treat  α (g, x) = 1,
by writing α (g, x) := α (g)(x). I will leave it to context to determine which of
these representations is intended. In either case, I will refer to 
α as the dynamical
state associated with state α ∈ . The state-separation condition (27.4) tells us
that α → α ∈ (E)G is injective, so if we wish, we can identify states with the
corresponding dynamical states.
The mapping  α ∈ (E)G is a (discrete) random probability measure on G, while

α : G × E → [0, 1] is a discrete Markov kernel on G × E. Thus, nothing we have
done so far takes us outside the range of classical probability theory. In fact, if G is
compact, there is a natural measure on G × E, namely the product of the normalized
Haar measure on G and the counting measure on E, such that
  & ' 

α (g, x)d(g, x) = 
α (g, x) dg = 1dg = 1.
G×E G x∈E G

In other words,for each state α ∈ , 

α (g, x) defines a probability density on G×E.
Moreover, as x∈E  α (g, x) = 1, the conditional density α (x|g) is exactly 
α (g) ∈
(E). Nevertheless, any symmetric pobabilistic model provides an example of this
scenario: simply chose a test E ∈ M(A), and let p(α, x) = α(x) and  α (g)(x) =
α(gx) for all α ∈ (A) and any x ∈ E. Conversely, as we will see, one can
reconstruct a symmetric probabilistic model A with E ∈ M(A) and  (A)
from the “classical” data described above, that is, the test E, the state space , the
symmetry group G and the probabilistic coupling p.

7 I prefer to write this as p(α, x) rather than as p(x|α) in order to make the covariance conditions

below come out more prettily.

608 A. Wilce

Definition 4.4 (Implementing a permutation) Let E, , G and p be as above,

and let σ be a permutation of the outcome-set E. Then we shall say that g ∈ G
implements σ iff p(αg, x) = p(α, σ x) for all α ∈  and all x ∈ E.
Our outcome-separation assumption (2.7.3) implies that if p(α, σ x) = p(α, τ x)
for all α ∈  and all x ∈ E, then σ x = τ x, i.e., σ = τ . Thus, if g ∈ G implements
a permutation σ ∈ S(E), it implements only one such permutation, which we may
denote by σ g .
Lemma 4.5 Let H be the set of all group elements g ∈ G implementing
permutations of E. Then H is a subgroup of G and σ : H → S(E), g → σ g ,
is a homomorphism.
Proof Let g, h ∈ H. Then for all α ∈  and all x ∈ E, we have

p(αgh, x) = p(αg, σ h x) = p(α, σ g σ h x).

Hence, gh ∈ H with σ gh = σ g σ h . We also have

p(αg −1 , x) = p(αg −1 , σ g σ −1 −1 −1
g x) = p(αg g, σ g x) = p(α, σ
−1
x)

so g −1 ∈ H . *
)
Thus, E carries a natural H action. In what follows, I will simplify the notation
by writing hx for σ h x, where h ∈ H and x ∈ E. It is natural to consider cases in
which E is transitive as an H-set, meaning that for every x, y ∈ E, there exits at
least one h ∈ H with hx = y. This motivates the following
Definition 4.6 (Dynamical classical models) A dynamical classical model
(DCM) consists of groups H ≤ G, a transitive left H -set E, a convex right G-
space , and a Markov kernel p :  × E → [0, 1] such that p(α, hx) = p(αh, x)
for all α ∈ , h ∈ H and x ∈ E.
Given a DCM as defined above, for each state α ∈  define 
α : G → (E) by


α (g)(x) = p(αg, x).

We shall say that a DCM is state-determining iff α →  α is injective, i.e., the

probability weights α (g) on (E), as g ranges over G, determine the state α.
Henceforth, I will assume that all DCMs are state-determining.
Any DCM can be reinterpreted as a symmetric dynamical-model in a routine
way, which we shall now describe. Let xo ∈ E and set

K := { g ∈ G | p(αg, xo ) = p(α, xo ) ∀α ∈  }.
27 Dynamical States and the Conventionality of (Non-) Classicality 609

Arguing in much the same way as above, it is easy to see that K, too, is a subgroup
of G. The following slightly extends a result from Wilce (2005); see also Wilce
(2009):
Theorem 4.7 Let E be a set acted upon transitively by a group H , let G be any
group with H ≤ G, and let K be any subgroup of G with K ≤ K and K ∩ H =
Hxo . Then there is a well-defined H -equivariant injection φ : E → X given by
φ(hxo ) = hK for all h ∈ H . Identifying E with φ(E) ⊆ X,
(a) M := { gφ(E) | g ∈ G } is a fully G-symmetric test space, and
(b) For every α ∈ ,

α (xg ) := α(g, xo )

is a well-defined probability weight on M; and
(c) The mapping α → α is a G-equivariant affine injection.

The proof is given in Appendix C.
Thus, the essentially classical picture presented above—a single test (or observ-
able) E, interacting probabilistically with a system with a state-space  and
symmetry group G by means of the function p :  × E → [0, 1]—can be
reinterpreted as a “non-classical” probabilistic model in which E appears as one
of many possible tests, provided that a sufficiently large set of permutations of E
can be implemented physically by the group G. What is more important for our
purposes, however, is the (even more) trivial converse: Any symmetric model A, and
in particular, any (finite-dimensional) quantum model, arises from this construction:
simply take H to be the stabilizer in G(A) of a chosen test E ∈ M(A) and K
to be the stabilizer in G(A) of some chosen outcome xo ∈ E, and we recover
(M(A), (A)) from Theorem 4.7.
To summarize: in the representation above, we have reinterpreted the states as, in
effect, random probability measures (or rather, weights) on a fixed test E, indexed
by the dynamical group G(A). Another way to view this is that each state determines
a family of trajectories in the simplex (E) of classical probability weights on E.
Given a state α ∈  plus the system’s actual dynamics, as specified by a choice of
one-parameter group g : R → G(A), we obtain a path  α t := 
α (gt ) in (E). In this
representation, notice,
(a) States are not viewed as probability weights on a non-classical test space, but
rather, specify how classical probabilities change over time, given the dynamics.
(b) In general, the trajectories in (E) arising from states and one-parameter
subgroups of G(A) are not governed by flows, that is, there is no one-parameter
group of affine mappings T s : (E) → (E) such that  α t+s = T s (α t ). In
particular, the observed evolution of probabilities on E is not Markov. In this
respect, it is the dynamical, rather than the probabilistic, structure of the DCM
that can be regarded as non-classical.
610 A. Wilce

It is also important to note that the representation of a symmetric probabilistic model

as a DCM is not in any usual sense a “hidden variables” one: it invokes no additional
structure, but is simply a different, and mathematically equivalent, way of viewing
the model (in particular, its states)—one that, I am urging, we might nevertheless
want to count as probabilistically “classical”: to whatever extent it departs from
classicality, it does so with respect to its dynamical, and not with respect to its
probabilistic, structure.

27.5 Composite Models, Entanglement and Locality

Thus far, we have been discussing probabilistic models largely in isolation, and
largely in abstraction from physics. Once we start to consider probabilistic models
in relation to one another, and as representing actual physical systems localized in
spacetime, e.g., particles (or laboratories), we encounter a host of new questions
regarding compound systems in which two or more component systems occupy
causally separate regions of spacetime. In particular, we unavoidably encounter the
concepts of entanglement and locality.
In this section, I briefly review how these notions unfold in the context of general
probabilistic theories (following Barnum and Wilce (2016), Foulis and Randall
(1981)), and how the various classical representations discussed above interact with
such composites. As we’ll see, entangled states arise naturally in this context. There
is a sense in which the representations in terms of classical extensions and in terms
of semiclassical covers are both obviously non-local, simply because in each case the
“classical” state spaces associated with a composite system AB is typically much
larger than the Cartesian product of those associated with A and B separately. Thus,
for instance, if S(A) stands for the set of dispersion-free probability weights on

M(A), then S(AB) is going to be much larger than S(A) × S(B). Similarly, unless
(A) or (B) is a simplex, (AB) will generally allow for entangled pure states,
which essentially just means that (AB)ext will be larger than (A)ext × (B)ext .
A more technical notion of “non-locality” in terms of hidden variables is discussed
below in Sect. 27.5.2, while the more delicate question of whether composites of
DCMs should be regarded as local or not is discussed in Sect. 27.5.4.

27.5.1 Composites of Probabilistic Models

Suppose A and B are probabilistic models. A joint probability weight on the test
spaces M(A) and M(B) is a function ω : X(A) × X(B) → [0, 1] that sums to
1 on each product test E × F , where E ∈ M(A) and F ∈ M(B). Certainly if α
and β are probability weights on A and B, we can form a joint probability weight
27 Dynamical States and the Conventionality of (Non-) Classicality 611

α ⊗ β on A and B, by the obvious recipe (α ⊗ β)(x, y) = α(x)β(y) for outcomes

x ∈ X(A) and y ∈ X(B). Note that joint probability weights can be understood as
probability weights on the test space

M(A) × M(B) := { E × F | E ∈ M(A), F ∈ M(B) }

consisting of all product tests.

Definition 5.1 (Non-signaling joint probability weights) A joint probability
weight ω on M(A) and M(B) is non-signaling iff its marginal (or reduced)
probability weights, given by

ω2 (y) := ω(E × {y}) and ω1 (x) := ω({x} × F )

are well-defined, i.e., independent of the choice of E ∈ M(A) and F ∈ M(B),

respectively.
To the extent to which we think of the tests E ∈ M(A) as representing the
physical actions performable on the system represented by A (in the sense that, to
perform a test is to do something to the system, and then observe a result), and
similarly for B, the non-signaling condition on ω tells us that no action performable
on A can have any statistically detectable influence on B, and vice versa. Thus, the
no-signaling condition is the locality condition appropriate to states qua probability
weights.
A non-signaling joint probability weight, having well-defined marginals, also has
well-defined conditional probability weights given by

ω(x, y) ω(x, y)
ω2|x (y) := and ω1|y (x) :=
ω2 (y) ω1 (x)

(with, say, the convention that both are zero if their denominators are). The marginal
states can be recovered as convex combinations of these conditional states, in a
version of the law of total probability:
 
ω1 = ω2 (y)ω1|y and ω2 = ω1 (x)ω2|x .
y∈F x∈E

We should certainly want these conditional and (hence) marginal states to belong to
the designated state-spaces of each component of a reasonable composite model.
Definition 5.2 Given probabilistic models A and B, let A ⊗ B denote the prob-
abilistic model with test space M(A ⊗ B) := M(A) × M(B) and state-space
(A ⊗ B) consisting of all non-signaling joint states ω on M(A) × M(B) having
conditional (and hence, marginal) states belonging to (A) and (B).
This is essentially the maximal (or injective) tensor product of (A) and (B)
(Barnum and Wilce 2016; Namioka and Phelps 1969). Since it does not allow for
612 A. Wilce

non-product outcomes, it is too small to be generally useful as a model of composite

systems, but it does afford an easy way to define something more general:
Definition 5.3 (Non-signaling composites (Barnum and Wilce 2016)) A (non-
signaling) composite of probabilistic models A and B is a model AB, together
with
(a) an outcome-preserving interpretation φ : A ⊗ B → AB, taking each pair of
outcomes x ∈ X(A) and y ∈ X(B) to a product outcome xy := φ(x, y),8 and
(b) a bi-affine mapping (A) × (B) → (AB) taking states α ∈ (A) and
β ∈ (B) to a product state α ⊗ β such that (α ⊗ β)(xy) = α(x)β(y) for all
x ∈ X(A), y ∈ X(B).
Examples 5.4 Consider quantum probabilistic models A(H) and A(K). If we take
AB to be A(H⊗K), then we have a natural map φ : X(H)×X(K) → X(H⊗K),
given by φ(x, y) = x ⊗ y. It is easy to check that this is an interpretation from
A(H) × A(K). We also have a natural bi-affine mapping (H) × (K) →
(H ⊗ K) given by, (α T , α W ) → α T ⊗W (where, recall, α T represents the
probability weight determined by the density operator T ). It is easy to see that these
mappings satisfy the conditions above.
If AB is a composite of models A and B as in Definition 5.3, every state ω ∈
(AB) pulls back along φ to a nonsignaling joint probability weight φ ∗ (ω)(x, y) =
ω(xy) on M(A) × M(B) with conditional and marginal weights belonging to
(A) and (B). In general, this joint probability weight does not determine the
state ω.
Definition 5.5 (Local tomography) A composite φ : A ⊗ B → AB of proba-
bilistic models A and B is locally tomographic iff the mapping φ ∗ : (AB) →
(A × B) is injective.
It is easy to show that, for finite-dimensional probabilistic models A and B,
AB is locally tomographic iff the mapping ⊗ : (A) × (B) → (AB)
extends to a linear isomorphism V(A) ⊗ V(B) V(AB). Mathematically, then,
local tomography is a great convenience. However, given its failure for real and
quaternionic QM, unless we are looking for an excuse to rule out these variants of
QM (and see, e.g., (Baez 2012) for reasons why we might not want to do so), it is
probably best to avoid taking local tomography as an axiom.
It is natural to ask that a probabilistic theory be closed under some rule of
composition, in the sense of Definition 5.3, that is consonant with the theory’s
categorical structure. A way to make this precise is to ask that a probabilistic
theory be a symmetric monoidal category. See Barnum and Wilce (2016) for further
discussion.

8 A more general definition would drop the requirement that φ be outcome-preserving, thus

allowing for the possibility that for some outcomes x and y, φ(x, y) =: xy might be a non-trivial
(and even possibly empty) event of AB. We will not need this generality here.
27 Dynamical States and the Conventionality of (Non-) Classicality 613

27.5.2 Entanglement

If AB is a 
composite of models A and B, then any convex combination of product
states, say i ti α i ⊗ β i or, more generally,

(α λ ⊗ β λ )dμ(λ)9


is said to be separable. Such states can (in principle) be prepared by randomly

selecting product states α λ and β λ for A and B, respectively, in a classically
correlated way. States not of this form are said to be entangled. The existence, and
the basic properties, of entangled states are rather generic features of probabilistic
models having non-simplex state spaces. In particular, we have the following
Lemma 5.6 Let ω be any non-signaling state on AB. Then
(a) If α ⊗ β is pure, then so are α and β;
(b) If either of the marginal states ω1 ∈ (A) or ω2 ∈ (B) is pure, then ω =
ω1 ⊗ ω2 ;
(c) Hence, if ω is entangled, then ω1 and ω2 are mixed.
Proof See Barnum and Wilce (2016) (and note that the arguments are virtually
identical to the familiar ones in the context of QM).10
When A and B are semiclassical, the test space M(A ⊗ B) defined in
Definition 5.2 is again semiclassical. In this situation, we have lots of dispersion-
free joint states. However:
Lemma 5.7 Let A and B be models with semiclassical test spaces M(A) and
M(B), and let ω ∈ Pr(M(A × B)). If ω is non-signaling and dispersion-free,
then ω = δ ⊗ γ where δ and γ are dispersion-free.
Proof Suppose ω is dispersion-free. Since ω is also non-signaling, it has well-
defined marginal states, which must obviously also be dispersion-free, hence, pure.
But by Lemma 5.6, a non-signaling state with pure marginals is the product of these
marginals. *
)
It follows that any average of non-signaling, dispersion-free states is separable.

9 Where  is a measurable space and μ is a probability measure thereon, and where the integral
exists in the obvious sense that
  
α λ ⊗ β λ )dμ(λ) (z) = (α λ ⊗ β λ )(z)dμ(λ).
 
10 As an historical note, both of these points were first noted in this generality, but without any
reference to entanglement, in a pioneering paper of Namioka and Phelps (1969) on tensor products
of compact convex sets. They were rediscovered, and connected with entanglement, by Kläy
(1988).
614 A. Wilce

27.5.3 Locality and Hidden Variables

In the literature, a (contextual) local hidden variable model for a bipartite probability
weight ω is usually taken to consist of a measurable space , a probability measure
μ thereon, and a pair of response functions pA and pB such that, such that for
all tests E ∈ M(A), F ∈ M(B) and all outcomes x ∈ E and y ∈ F ∈
M(B), p(x|E, λ) is the probability of x given the parameter λ and the choice of
measurement E, and similarly for p(y|F, λ). These are required to satisfy

ω(x, y) = pA (x|E, λ)p(y|F, λ)dμ(λ)


so that we can interpret the joint probability ω(x, y) as resulting from a classical
correlation (given by μ) together with the local response functions pA and pB .
As discussed in Sect. 27.3.3, it is straightforward to reinterpret the functions pA
and pB in terms of probability weights on M(A)  
and M(B), respectively, by
writing pA (x|E, λ) as pA,λ (x, E) and pB (y|F, λ) as pB,λ (y, F ) (emphasizing that
λ can be treated as merely an index). We then have
 
ω(x, y) = pA,λ ⊗ pB,λ dμ(λ) ((x, E), (y, F )),


independently of E and F . In other words, ω has a local HV model if, and only if,
it is separable, in the integral sense, when regarded as a joint probability weight on
 ⊗ B) (using the composite of Definition 5.2).
M(A
Remark A first reaction to this observation might be that it must be wrong, as it’s
well-known that there exist entangled Werner states that nevertheless have local
HV representations (Werner 1989). The subtlety here (such as it is) resides in the
fact that whether a state is separable or entangled depends on what “local” state-
spaces are in play. Here, we are expressing ω as a weighted average of products
of “contextual” states, whereas in standard discussions of quantum entanglement,
a state is separable iff it can be expressed as a weighted average of products of
quantum (in particular, non-contextual) states of the component systems.

27.5.4 Composites of Dynamical Models

Since dynamical classical models are simply another way of looking at symmetric
probabilistic models, in principle they can be composed in the same way as the latter.
I have discussed this in some technical detail elsewhere (Wilce 2009). Without going
into such detail, I want to make a few remarks on the sense in which composites of
DCMs can be regarded as local. Suppose, then, that

A = ((A), G(A), H (A), E(A), pA ) and B = ((B), G(B), H (B), E(B), pB )

27 Dynamical States and the Conventionality of (Non-) Classicality 615

are two DCMs as defined in Sect. 27.4.2, and suppose that AB is a DCM serving as
a composite of these. This will require, at a minimum that E(AB) = E(A) × E(B),
and that we have functions

⊗ : (A) × (B) → (AB) and ⊗ : G(A) × G(B) → G(AB)

with the former bi-affine and the latter a group homomorphism, such that

(G(A) ⊗ G(B)) ∩ H (AB) ≤ H (A) ⊗ H (B)

and

(α ⊗ β)(g ⊗ h) = αg ⊗ βh

for all (α, β) ∈ (A) × (B) and all (g, h) ∈ G(A) × G(B). Additional conditions
are necessary to ensure that a non-signaling condition is satisfied, but this is enough
for present purposes.
Examples 5.8
(a) By the minimal classical dynamical model associated with a finite set E, I mean
the DCM A(E) with (A(E)) = (E), G(A(E)) = H (A(E)) = S(E) (the
symmetric group of all bijections E → E) and pA(E) (α, x) = α(x). Now given
two finite sets, then the minimal DCM A(E × F ) is a composite of A(E) and
A(F ) with the maps ⊗ : (E)×(F ) → (E ×F ) and ⊗ : S(E)×S(F ) →
S(E × F ) the obvious ones, that is,

(α ⊗ β)(x, y) = α(x)β(y) and (g, h)(x, y) = (gx, hy)

for all (x, y) ∈ E × F , (α, β) ∈ (E) × (F ) and (g, h) ∈ S(E) × S(F ).
(b) Let A and B be quantum-mechanical systems associated with Hilbert spaces
HA and HB . Fixing orthonormal bases E ∈ M(HA ) and F ∈ M(HB ),
we can regard A and B as DCMs with (A) and (B) the spaces of density
operators on HA and HB , respectively, G(A) = U (HA ) and G(B) = U (HB ),
and with

pA (ρ, x) = Tr(ρpx ) = ρx, x

and similarly for pB . Identifying E × F with the product basis E ⊗ F in

M(HA ⊗ HB ), we can regard the quantum system associated with HA ⊗ HB
as a DCM in the same way. In this case the maps ⊗ referred to above are the
obvious ones: given unitaries U ∈ U (HA ) and V ∈ U (HB ), we have a untary
U ⊗ V on HA ⊗ HB , and if ρ A and ρ B are density operators on HA and HB
respectively, the ρ A ⊗ ρ B is a density operator on HA ⊗ HB .
As discussed above, the prevailing technical definition of a local probabilistic
theory is one in which composite systems admit local hidden-variable models
(ideally, in a functorial way, though this point is rarely discussed). More generally,
one would like to say that a probabilistic theory is local iff composite systems admit
616 A. Wilce

“local” classical representations, in some well-defined (but reasonably general)

sense. As I’ve mentioned, the classical extensions and the classical representations
associated with semiclassical covers, discussed in Sects. 27.3.2 and 27.3.3, are
certainly not local in any reasonable sense. What about composites of DCMs? In
what sense are, or aren’t, these to be regarded as local classical representations
of symmetric probabilistic models? As the second example above illustrates, the
state space (AB) will in general be significantly larger (A) × (B) in the
sense that there will be entangled states. In this weak sense, such a composite will
generally be “non-local”. On the other hand, it is not entirely clear how to discuss the
locality of a composite of dynamical classical models as a classical representation,
since it involves only a single, classical measurement E(A) × E(B), and invokes
no hidden variables, non-local or otherwise—unless, of course, we wish to regard
the symmetries in G(AB) as “hidden variables”. In that case, the “non-locality”
presumably rests in the fact that the group G(AB) will typically be a great deal
larger than G(A) × G(B). This is true, of course, for quantum models, where the
unitary group U (HA ⊗ HB ) is larger than U (HA ) ⊗ U (HB ). But it is also true for
minimal classical dynamical models, since for sets E and F , the symmetric group
S(E × F ) is larger than S(E) × S(F ).

27.6 Conclusion

Many of the observations collected above are well known, at least folklorically. My
aim in bringing them together, in the particular way that I have, has been to draw
attention to some points that I believe follow from these observations collectively,
and that I believe are somewhat less widely appreciated than they might be:
(a) Classicality in the strict (that is, unrestricted) Kolmogorovian sense, is man-
ifestly a contingent matter, since it is not a point of logical necessity that
all experiments should be compatible. To the extent that probability theory is
meant to be a completely general and a priori study of reasoning in the face of
uncertainty or chance, it isn’t Kolmogorovian, nor is it quantum-mechanical:
it is, rather, the study of what I’ve here called probabilistic theories generally
(whether formulated as I have done here, or in some roughly similar way).
(b) Within such a framework, what we (ought to) mean by saying that a probabilis-
tic model or probabilistic theory is “classical” is not entirely obvious. For some
values of “classical”, every probabilistic theory is, or can be represented as, a
classical one, at the cost of contextuality and, where we are dealing with a non-
signaling theory, also of locality. Moreover, for certain kinds of theories (fully
symmetric ones), there is a broad sense of “classical” in which such a theory
simply is classical, without any need to invoke hidden variables, contextual,
non-local or otherwise. Or at least, this is true for the theory’s probabilistic
structure: any departure from “classicality” is dynamical. Finite-dimensional
quantum mechanics is such a theory.
27 Dynamical States and the Conventionality of (Non-) Classicality 617

(c) The selection of a class of probabilistic models—a probabilistic theory—must

rest on some contingent assumptions about the nature of the entities one seeks to
model. Quantum mechanics, as a probabilistic theory, is contingent in this way;
so is (unrestricted, Kolmogorovian) classical probability theory, and the two are
not the same. In this limited and unremarkable sense, quantum mechanics is (of
course) a non-classical probabilistic theory.
I don’t know how much, if any, of this Pitowsky would have endorsed. Regarding
(b) at least, he was very clear:
We can always avoid the radical view of probability by adopting a non-local, contextual
hidden variables theory such as Bohm’s. But then I believe, the philosophical point is
missed. It is like taking Steven Weinberg’s position on space-time in general relativity:
There is no non-flat Riemannian geometry, only a gravitational field defined on a flat space-
time that appears as if it gives rise to geometry . . . I think that Weinberg’s point and also
Bohm’s theory are justified only to the extent that they yield new discoveries in physics (as
Weinberg certainly hoped). So far they haven’t.

I have no issue to take with this, as far as it goes: we are always free to reject
a classical representation of a probabilistic theory by objecting to the additional
“classical” structure as “non-physical”. But I would also want to say that doing
so makes the theory a physical theory, and not a theory of probability. Or, to put
it differently, whether one wants to take, for example, a set of contextual hidden
variables, or a privileged observable, seriously in formulating a given physical
theory is a pragmatic question about the best (the most elegant, the most convenient,
the most fruitful) way to formulate mathematical models of a particular sort of
physical situation. It is also (in particular!) a metaphysical question about what
ontology one wants to admit in framing such a model. Such decisions are important;
but they are not decisions about what kind of probability theory to use.
Pitowsky also says (emphasis mine):
In this paper, all we have discussed is the Hilbert space formalism. I have argued that it is
a new kind of probability theory that is quite devoid of physical content, save perhaps the
indeterminacy principle which is built into axiom H4. Within this formal context there is no
explication of what a measurement is, only the identification of “observables” as Hermitian
operators. In this respect the Hilbert space formalism is really just a syntax which represents
the set of all possible outcomes, of all possible measurements.

Here, of course, I disagree. The “syntax” of quantum probability theory is very

definitely not devoid of physical content. Reconstructions of QM from various sets
of postulates, ranging from the Piron-Amemiya-Araki-Solèr reconstruction cited
by Pitowsky (see Holland 1995), to the various more recent finite-dimensional
reconstructions (e.g., Mueller et al. 2014; D’Ariano et al. 2011; Dakič and Brukner
2011; Masanes and Müller 2011; Wilce 2012) have helped us understand the phys-
ical content of quantum theory more clearly, by isolating operationally meaningful
principles that dictate that syntax. These principles invariably include some that are
manifestly contingent, at least to the same degree that the classical co-measurability
principle is contingent.
618 A. Wilce

The final lines of Pitowsky’s paper crystalize what I find problematic in his
proposal. (Again, the emphasis is mine.)
However, there is a structure to the set of events. Not only does each and every type of
measurement yield a systematic outcome; but also the set of all possible outcomes of all
measurements — including those that have been realized by an actual recording — hang
together tightly in the structure of L(H). This is the quantum mechanical structure of
reality.

Taken in the context of Pitowsky’s claim that QM represents a new, non-classical

theory of probability, this suggests a certain naturalism regarding probability theory
(echoing von Neumann/Birkhoff’s naturalism regarding logic): that there is a true
probability theory, which is the one that works best, and most broadly, to describe
the world we actually live in. If the world really is, at every scale and in every
way, described by some version of quantum theory, so that there just are no actual
events that can’t be described as effects in (say) some grand von Neumann algebra,
then quantum probability theory is this true, correct probability theory, of which
classical probability theory is simply a limiting or special case (where h̄ tends to
zero, or where we restrict attention to a set of commuting observables).
But this view seems to me misleading: not because naturalism per se is wrong
(even about mathematics), but because it forgets that probability theory has a
methodological as well as a metaphysical importance: it is not simply part of our
scientific description of the world, but is part of the framework within which we
think about, criticize, and (often) revise that description. In order for it to play this
role, it can not be tied to any particular physical theory. This is why I think it is vital
to maintain the distinction between probability theory and the various probabilistic
theories that it studies. There’s no such thing as a probability theory, classical or
otherwise: there is only probability theory. There can be a correct probabilistic
theory of the world—and perhaps it’s quantum.

Appendix A Common Refinements

Let A be a probabilistic model, and let E, F ∈ M(A). Recall that E refines F , or

that F is a coarsening of E, if there exists a surjection f : E → F such that, for
every y ∈ F and every state α ∈ ,

α(y) = α(f −1 (y)).

When this is so, we write E / F , and say that f is a coarsening map.

Definition A.1 A set  of probability weights on a test space M separates
compatible events iff, for every test E ∈ M and any pair of distinct events a, b ⊆ E,
there exists a weight α ∈  with α(a) = α(b).
27 Dynamical States and the Conventionality of (Non-) Classicality 619

Note that if A is unital, (A) separates compatible events. Indeed, it is sufficient

that, for every x ∈ E, there exist a state α ∈ (A) with α(x) > 1/2.
Lemma A.2 Let (A) separate compatible events. Then
(a) There exists at most one coarsening map f : E → F ;
(b) If there exist coarsening maps f : E → F and g : F → E, then f and g are
bijections and g = f −1 ;
(c) If f : G → E and g : G → F are coarsening maps and x ∈ E ∩ F , then
f −1 (x) = g −1 (x) ⊆ G.
Proof (a) If f, g : E → F are coarsening maps, where E / F in M(A), then
for every y ∈ F we have α(f −1 (y)) = α(y) = α(g −1 (y)) for every α ∈ (A).
Hence, f −1 (y) = g −1 (y) for every y ∈ F , whence, f = g. Now (b) follows, since
the composition of coarsening maps is a coarsening map. For (c), observe that if
x ∈ E ∩ F , then for every state α we have α(f −1 (x)) = α(x) = α(g −1 (x)). Since
 separates compatible events, f −1 (x) = g −1 (x). *
)
From now on, assume (A) separates compatible events. We will write fE,F for
the unique coarsening map f : E → F , if one exists.
Suppose now that E, F ∈ M(A) have a common refinement, that is, that there
exists a test G ∈ M(A) with G / E and G / F . Then we have a natural surjection
φ : G → E × F , namely

φ(z) = (fG,F (z), fG,E (z)).

If α ∈ , then we have a probability weight on E × F given by

φ ∗ (α) := α ◦ φ −1 .

This assigns to (x, y) ∈ E × F the probability

−1 −1
α(φ −1 (x, y)) = α(ax ∩ by ), where ax = fG,E (x) and by = fG,F (y).
 
It is easy to check that y∈F φ ∗ (α)(x, y) = α(x) and x∈E φ ∗ (α)(x, y) = α(y),
i.e., φ ∗ (α) has the “right” marginals to explain the probabilities that α assigns to E
and F . In this sense, G can be regarded as a joint measurement of E and F .
Definition A.3 A is a refinement ideal iff every pair of tests in M(A) has a
common refinement.
In other words, A is a refinement ideal iff the preordered set (M, /) is
downwardly directed. Define S ⊆ E∈M E to be the set

S = { x = (xE ) ∈ E∈M(A) | ∀E / F fE,F (xE ) = xF }, (27.5)

i.e., the inverse limit of M, regarded as a small category under coarsening maps.
As long as M is locally finite, one can show that this is non-empty (a consequence
620 A. Wilce

of the compactness of E∈M E). For each test E ∈ M(A), define fE : S → E by

fE (x) = xE , and for any x0 ∈ X(A), let

[x] := { x ∈ S | ∀E ∈ M(A) x0 ∈ E ⇒ xE = x0 }.

Note that, by Lemma A.2, if x0 ∈ E ∩ F , then for every x ∈ S, xE  = x iff xF = x0 ;

hence, [x0 ] = ∅. If a ∈ Ev(A), let [[a]] := {[x]|x ∈ a}. Notice that [[a]] = fE−1 (a)
where E is any test with a ⊆ E.
Recall that, if E and F are partitions of a set S, E is said to refine F iff for every
y ∈ F , there is some a ⊆ E with y = ∪a. I will write E 0 F to indicate this.
If M(A) is a collection of partitions of a set S such that every pair of partitions
in M(A) has a common refinement in this sense, then I will say that M(A) is a
refinement ideal of partitions.
Lemma A.4 If A is a locally finite refinement ideal, then with S the projective limit
of M(A) in (27.5), For every E ∈ M(A), [[E]] is a partition of S, and if E / F
then [[E]] 0 [[F ]]. Hence, [[M(A)]] := {[[E]]|E ∈ M(A)} is a refinement ideal
of partitions. Moreover, the mapping φ : x → [x] defines an isomorphism of test
spaces from M(A) to [[M(A)]].
Proof That [[E]] is a partition of S is clear from the definitions. Let y ∈ F , and set
−1
a = fE,F (y) ⊆ E. Thus, [[a]] ⊆ [[E]]. Now

[[a]] = fE−1 (a) = fE−1 ◦ fE,F
−1
(x) = (fE,F ◦ fE )−1 (x) = fF−1 (y) = [y].

Thus, every cell in the partition [[F ]] is a union of cells of [[E]], i.e., E 0 F . Since
every pair of tests in M(A) have a common refinement with respect to (A), it
follows that [[M(A)]] is a refinement ideal of partitions. It is clear that φ : x →
[x] is an outcome-preserving, positive interpretation from M(A) onto [[M(A)]].
It remains to show it’s injective. Let x, y ∈ X(A) with [x] = [y]. Supposing that
−1 −1
x ∈ E ∈ M(A) and y ∈ F ∈ M(A), let G / E, F . Then fG,E (x) = fG,F (y), so
for all α ∈ (A), we have
−1 −1
α(x) = α(fG,E (x)) = α(fG,F (y)) = α(y).

By our standing assumption that states in (A) separate outcomes in X(A), x = y.

*
)
Now let M be a test space of partitions of a set S, and let M be a refinement
ideal with respect to ordinary refinement of partitions.
Define φ : Ev(M) → P(S) by  a = ∪a for all a ∈ Ev(M). Note here that an
event a is a subset of a finite partition of S, that is, a finite, pairwise disjoint set of
non-empty subsets of S. Now let

 := { 
a | a ∈ Ev(M) }.
27 Dynamical States and the Conventionality of (Non-) Classicality 621

Lemma A.5  is an algebra of sets on S.

Proof It is easy to see that if a, b ⊆ E ∈ M,  a ∩
b = a ∩ b, a∪b = a ∪ b, and
 
a = E  a. Thus, the set E := {
c a |a ⊆ E} is a subalgebra of P(S). If G 0 E,
then the definition of refinement tells us that E ⊆ G . Hence, if M is a refinement
ideal, {E |E ∈ M} is a directed family of subalgebras of P(S) under inclusion,
whence, its union is also a subalgebra. *
)
Now let A be a refinement ideal in the sense of probabilistic models, i.e., M(A)
is a refinement ideal with respect to (A). For every α ∈ (A), define ∀  a ∈ ,

 a ) = α(a).
α (

Lemma A.6 
α is well-defined, and a finitely-additive probability measure on .
Proof Let a ⊆ E ∈ M and b ⊆ F ∈ M. We want to show that if  a = 
b,
then α(a) = α(b). Let G refine both E and F . There are canonical surjections
e : G → E and f : G → F , namely e(z) = x where x is the unique cell of E
containing z ∈ G, and similarly for f . Let a1 = e−1 (a) and b1 = f −1 (b). Then
a1 = a = b = b1 . Since  is injective on P(G), a1 = b1 . Since A is a refinement
ideal with respect to (A), α(a) = α(a1 ) = α(b1 ) = α(b).
This shows that  α is well defined. To see that it’s additive, let a ⊆ E ∈ M(A),
b ⊆ F ∈ M(A), with  a ∩b = ∅. Choosing G a common refinement of E and F ,
we have a1 , b1 ⊆ G with  a1 = a and b1 =  a1 ∩ 
b, so  b1 = ∅, whence, a1 ∩ b1 = ∅.
It follows that

 a ∪
α ( α (a
b) =  1 ∪ a2 ) = α(a1 ∪ a2 ) = α(a1 ) + α(a2 )

=
α ( α (
a1 ) +  b1 ) =  α (
a) + 
α ( b).

*
)
Proposition A.7 If A is a locally finite refinement ideal and (A) separates
compatible events, then there is a measurable space (S, ) and an embedding
M(A) → D(S, ) such that every probability weight in  extends to a finitely-
additive probability measure on (S, ). Conversely, if A is unital and locally finite
and admits such an embedding, then A embeds in a refinement ideal without loss of
states.
Proof The forward implication follows from Lemmas A.4, A.5 and A.6, while the
reverse implication is more or less obvious. *
)
To this extent, then, the existence of common refinements is the key classical
postulate.
622 A. Wilce

Appendix B Semiclassical Test Spaces

Recall that M is semiclassical iff distinct tests in M are disjoint. Evidently, such a
test space has a wealth of dispersion-free states, as we can simply choose an element
xE from each test E ∈ M and set δ(x) = 1 if x = xE for the unique test containing
x, and 0 otherwise. In fact, the dispersion-free states are exactly the pure states:
Lemma B.1 Let Ki be an indexed family of convex sets, and let α = (α i ) ∈ K =
i∈I Ki . Then α is pure iff each α i is pure.

Proof Suppose that for some j ∈ I , α j is not pure. Then there exist distinct points
β j , γ j ∈ Kj with α j = tβ j + (1 − t)γ j for some t ∈ (0, 1). Define 
β, 
γ ∈ K by
setting
 
 α i i = j α i i = j
βi = γi =
and 
βj i = j γj i = j

Then we have t 
β + (1 − t)
γ = α, so α is not pure either. The converse is clear. )
*
Corollary B.2 If M is semiclassical, then Pr(M)ext = Pr(M)df .
Proof Since M is semiclassical, Pr(M) E∈M (E). *
)
Let S(M) denote the set of dispersion-free states on M. It is easy to see that
S(M) is closed, and hence compact, as a subset of [0, 1]X . By (S), I mean the
simplex of Borel probability measures on S with respect to this topology.

Appendix C Constructing Fully G-Symmetric Models

This appendix collects more technical material on the construction of fully symmet-
ric models. In particular, we prove Theorem 4.7, which for convenience we restate
below as Proposition C.2.
As discussed in Sect. 27.4.2, we are given a set E (which we wish to regard as
the outcome-set of an experiment), and a group H acting transitively on E. We are
also given a set  of physical states (those of the system to which the experiment
pertains), and a function

p :  × E → [0, 1]

assigning a probability weight 

α = p(α, ·) on E to each state α ∈ . We are also
given a group G of “physical symmetries” acting on  on the right, plus a subgroup
H of G acting on E (on the left), in such a way that

p(αh, x) = p(α, hx)

27 Dynamical States and the Conventionality of (Non-) Classicality 623

for all h ∈ H . For each α ∈ , we write  α both for the mapping G → (E)
given by α (g)(x) = p(αg, x), and for the mapping G × E → [0, 1] given by

α (g, x) = 
α (g)(x), whichever is more convenient. Thus, α3 g1 (g2 ) = 
α (g1 g2 ) for all
g1 , g2 ∈ G, and 
α (gh)(x) =  α (g)(hx) for all g ∈ G, h ∈ H and x ∈ E. We assume
that the mapping α →  α is injective, so that α is determined by the function  α.
Given this data, choose and fix an outcome xo ∈ E, and let

K := {k ∈ G|∀α ∈  p(αk, xo ) = p(α, xo )}. (27.6)

α (gk)(xo ) = 
Equivalently, k ∈ K iff  α (g)(xo ). Notice that the stabilizer Ho of xo in
H is a subset of K; in particular, K is nonempty.
Lemma C.1 K ≤ G, and K ∩ H = Ho .
Proof Let k, k  ∈ K. Then for all g ∈ G and α ∈ , we have α(gkk  , xo ) =
α(gk, xo ) = α(g, xo ) so kk  ∈ K; also

α (gk −1 , xo ) = 
 α ((gk −1 )k, xo ) = 
α (g, xo )

so k −1 ∈ K. Thus, K is a subgroup of G. For the second statement, h ∈ K ∩ H

implies that, for all α and all g,


α (g, xo ) = 
α (gh, xo ) = 
α (g, hxo ).

Since the functions α (g) separate points of E, hxo = xo , so h ∈ Ho On the other

hand, if h ∈ Ho , then α(gh, xo ) = α(g, hxo ) = α(g, xo ), so h ∈ K. *
)
We now prove Theorem 4.5, showing that any choice of a subgroup K between
Ho and K generates a fully G-symmetric test space, having  as its state space, in
a canonical manner. For convenience, we restate this result here:
Proposition C.2 With notation as above, suppose H acts transitively on E. Let K
be any subgroup of G with Ho ≤ K ≤ K. Set X := G/K, and let xg := gK ∈ X
for all g ∈ G. Then there is a well-defined H -equivariant injection φ : E → X
given by φ(hxo ) = xh for all h ∈ H . Moreover, identifying E with φ(E) ⊆ X,
(a) M := {gφ(E)|g ∈ G} is a G-symmetric test space;
(b) For every α ∈ , α (xg ) := 
α (g, xo ) is a well-defined probability weight on M;
(c) The mapping α  → α is a G-equivariant affine injection, where G acts on

(E)G on the right.
Proof For convenience, let us write xg for the left coset gK in X = G/K. Then the
mapping φ : E → X above is given by

φ : hxo → xh .

To see this is well-defined, let h, h ∈ H with hxo = h xo : then h−1 h ∈ Ho ≤ K,

so that xh = xh . Conversely, if xh = xh , we have h−1 h ∈ So whence hxo = h xo ,
624 A. Wilce

and φ is injective. If x = hxo , then φ(h x) = h hK = h xh = hφ(x), so φ is

H -equivariant. This disposes of (a). Now let M = { g E  | g ∈ G }. Since X is a
G-set, this gives us a G test space—a symmetric one, as G acts transitively on M
and H acts transitively on φ(E) ∈ M. For α ∈ , set α (xg ) = α(g, xo ). This is
 K, and thus
well-defined, since if xg = xg  , then g  = gk for some k ∈

α (xg  ) = α (xgk ) = 
α (gk, xo ) = 
α (g, xo ) = α (xg ).
  
To see that α is a probability weight on M, for each x ∈ E, choose hx ∈ H with

hx xo = x (recalling here that H acts transitively on E). Then

α (gxhx ) = 
α (gφ(x)) =  α (xghx ) = 
α (ghx , xo ) = 
α (g, hx xo ) = 
α (g, x).

Thus,
 
α (gφ(x)) = 
α (g, x) = 1.
x∈E
 x∈E

The mapping α → α is obviously affine, and is injective by our assumption that

α →   see that it is equivariant, note that for all g, l ∈ G and α ∈ ,
α is injective. To

(αg)(xl ) = (5
α g)(l, xo ) = 
α (gl, xo ) = α (glxo ) = α (g, xl ) = α g(xl ).
4   
*
)
Thus, (M, ), where  = {α |α ∈ }, is a G-symmetric model with state space
 .
isomorphic to  

Remarks
(1) If Ho ≤ K ≤ K  ≤ K, we obtain a G-equivariant, surjective outcome-
preserving interpretation

φ K  ,K : MK → MK 

given by φ K  ,K (gK) = gK  for all g ∈ G. So taking K = Ho gives in this

sense the least constrained symmetric G-test space containing E (in such a way
as to extend the action of H on E), while K = K gives the most constrained.
Note all such test spaces have the same rank, namely |E|, provided that the set
of probability weights α (g), with α ranging over  and g ranging over G, is
large enough to separate points of E.
(2) Accordingly, all probability weights on MK  pull back to probability weights
on MK via φ ∗K,K  , which is an injective G-equivariant affine mapping. Writing
K := Pr(MK ) for the convex set of all probability weights on MK , we have
 ⊆ K (where we identify  with its image ). Thus, in particular, we can

replace  with Pr(MK ) to obtain a larger state space with respect to which the
same construction works.
27 Dynamical States and the Conventionality of (Non-) Classicality 625

Given the H -set E, we can obtain the initial data for this construction as follows.
Let G be any group containing H as a subgroup, and let

G ×H E = (G × E)/ ∼

where ∼ is the equivalence relation defined by (g, x) ∼ (h, y) iff there exists some
s ∈ H with h = gs and sy = x. Letting [g, x] denote the equivalence class of
(g, x), we have

[g, x] = [gs, s −1 x] i.e., [gs, x] = [g, sx]

for all g ∈ G, x ∈ E and s ∈ H . Note that G ×H E is a G-set, with action defined

by

g[h, x] = [gh, x]

which is well-defined because [ghs, s −1 x] = [gh, x] for all s ∈ H . Now let

[g, E] := {[g, x]|x ∈ E} :

then if [h, x] ∈ [g, E], we have [h, x] = [g, y] for some y ∈ E, whence, there
is some s ∈ H with h = gs and sy = x. Then for all z ∈ E, say z = s  x, we
have [h, z] = [gs, s  x] = [g, ss  x] ∈ [g, E]. That is, [h, E] ⊆ [g, E]. By the same
token, [g, E] ⊆ [h, E]. In other words, the sets [g, E] partition G ×H E. With this
observation, it is easy to prove the following
Lemma C.3 With notation as in Proposition C.2, XHo G ×Ho E, and MHo
{[g, E]|g ∈ E}, a semiclassical test space, independent of .
Proof Let φ : G × E → G/Ho = XHo be given by

φ(g, sxo ) = gsHo .

for all g ∈ G, s ∈ H .
To see that this is well-defined, note that if s1 xo = s2 xo , then s2−1 s1 := s ∈ Ho ,
and gs1 Ho = gs2 sHo = gs2 Ho . Next, observe that φ(g1 , s1 xo ) = φ(g2 , s2 xo )
iff g1 s1 Ho = g2 s2 Ho iff g2 s2 = g1 s1 s for some s ∈ Ho , whence, [g2 , s2 xo ] =
[g2 s2 , xo ] = [g1 s1 s, xo ] = [g1 , s1 xo ]. Thus, passing to the quotient set G ×Ho E
gives us a bijection. Since {[g, E]|g ∈ G} is a partition of G ×Ho E, we have a
semiclassical test space, as advertised. *
)
Notice that the stabilizer of [e, E] in G is exactly H0 . Now choosing any G-
invariant set of probability weights on MHo —say, the orbit of any given probability
weight—we have a natural G-equivariant injection  → (E)G given by

α → 
α: 
α (g)(x) = α[g, x]
626 A. Wilce

for all α ∈ . The construction now proceeds as above for any choice of subgroup
K of G with Ho ≤ K ≤ K, where K is determined by  as in Equation (27.6). We
can regard K as parameterizing the possible fully G-symmetric models containing
E as a test, and having dynamical group G. If K = Ho , the stabilizer of xo in H ,
then M is semiclassical. Larger choices of K further constrain the structure of M,
enforcing outcome-identifications between tests that, in turn, constrain any further
enlargement of the state space that we might contemplate.

References

Alfsen, E. (1971). Convex sets and boundary integrals. Heidelberg/Berlin: Springer.

Bacciagaluppi, G. (2004). Classical extensions, classical representations, and Bayesian updating
in quantum mechanics. arXiv:quant-ph/0403055v1.
Baez, J. (2012). Division algebras and quantum theory. Foundations of Physics, 42, 819–855.
arXiv:1101.5690.
Barnum, H., & Wilce, A. (2016). Post-classical probability theory. In G. Chiribella & R. Spekkens
(Eds.), Quantum theory: Informational foundations and foils. Dordrecht: Springer.
Beltrametti, E., & Bugajski, S. (1995). A classical extension of quantum mechanics. Journal of
Physics A, 28, 3329.
Bub, J., & Pitowsky, I. (2012). Two dogmas about quantum mechanics. In S. Saunders, J. Barrett,
A. Kent, & D. Wallace (Eds.), Many worlds? Everett, quantum theory and reality (chapter 14,
pp. 433–459). Oxford. arXiv:0712.4258.
Dakič, B., & Brukner, C. (2011). Quantum theory and beyond: Is entanglement special? In
H. Halvorson (Ed.), Deep beauty: Understanding the quantum world through mathematical
innovation. Cambridge University Press. arXiv:0911.0695.
D’Ariano, M., Chiribella, G., & Perinotti, P. (2011). Informational derivation of quantum theory.
Physical Review A, 84, 012311.
Foulis, D. J., & Randall, C. H. (1978). Manuals, morphisms and quantum mechanics. In A. R.
Marlowe (Ed.), Mathematical foundations of quantum theory. New York: Academic Press.
Foulis, D. J., & Randall, C. H. (1981). Empirical logic and tensor products. In H. Neumann (Ed.),
Interpretations and foundations of quantum mechanics. Mannheim: B.I.-Wissenshaftsverlag.
Hardy, L. (2001). Quantum theory from five reasonable axioms. arXiv:quant-ph/0101012.
Hellwig and Singer. (1991). “classical” in terms of general statistical models. In V. Dodonov &
V. Man’ko (Eds.), Group theoretical methods in physics (Lecture notes in physics, Vol. 382).
Berlin/Heidelberg: Springer.
Hellwig, K.-E., Bush, S., & Stulpe, W. (1993). On classical representations of finite-dimensional
quantum mechanics. International Journal of Theoretical Physics, 32, 399.
Holevo, A. (1982). Probabilistic and statistical aspects of quantum theory. Amsterdam: North
Holland.
Holland, S. (1995). Orthomodularity in infinite dimensions: a theorem of m. solèr. Bulletin of the
American Mathematical Society, 32, 205–234. arXiv:math/9504224.
Kläy, M. (1988). Einstein-Podolsky-Rosen experiments: the structure of the probability space I, II.
Foundations of Physics Letters, 1, 205–244.
Lindsay, B. (1995). Mixture models: Theory, geometry and applications (NSF-CBMS regional
conference series in probability and statistics, Vol. 5). Hayward: Institute of Mathematical
Statistics/American Statistical Association.
Masanes, L., & Müller, M. (2011). A derivation of quantum theory from physical requirements.
New Journal of Physics, 13. arXiv:1004.1483.
27 Dynamical States and the Conventionality of (Non-) Classicality 627

Mueller, M., Barnum, H., & Ududec, C. (2014). Higher-order interference and single-system
postulates characterizing quantum theory. New Journal of Physics, 16. arXiv:1403.4147.
Namioka, I., & Phelps, R. (1969). Tensor products of compact convex sets. Pacific Journal of
Mathematics, 31, 469–480.
Pitowsky, I. (2005). Quantum mechanics as a theory of probability. In W. Demopolous (Ed.),
Physical theory and its interpretation: Essays in honor of Jeffrey Bub (Western Ontario series
in philosophy of science). Kluwer. arXiv:0510095.
Popescu, S., & Rohrlich, D. (1994). Quantum nonlocality as an axiom. Foundations of Physics,
24, 379–285.
Spekkens, R. (2005). Contextuality for preparations, transformations and unsharp measurements.
Physical Review A, 71. arXiv:quant-ph/0406166.
Varadarajan, V. S. (1985). The geometry of quantum theory (2nd ed.). New York: Springer.
Werner, R. (1989). Quantum states with Einstein-Podolsky-Rosen correlations admitting a hidden-
variable model. Phys. Rev. A, 40, 4277.
Wilce, A. (2005). Symmetry and topology in quantum logic. International Journal of Theoretical
Physics, 44, 2303–2316.
Wilce, A. (2009). Symmetry and composition in generalized probabilistic theories.
arXiv:0910.1527.
Wilce, A. (2012). Conjugates, filters and quantum mechanics. arXiv:1206.2897.
Wilce, A. (2017). Quantum logic and probability theory. In Stanford encyclopedia of philosophy.
https://plato.stanford.edu/archives/spr2017/entries/qt-quantlog/.